Support whisper models (#238)

2023-08-07 12:34:18 +08:00
parent 64efbd82af
commit 45b9d4ab37
39 changed files with 1836 additions and 52 deletions
--- a/.github/workflows/export-whisper-to-onnx.yaml
+++ b/.github/workflows/export-whisper-to-onnx.yaml
@@ -0,0 +1,63 @@
+name: export-whisper-to-onnx
+
+on:
+  workflow_dispatch:
+
+concurrency:
+  group: release-whisper-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  release-whisper-models:
+    if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
+    name: ${{ matrix.model }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [macos-latest]
+        model: ["tiny.en", "base.en", "small.en", "medium.en"]
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Install dependencies
+        shell: bash
+        run: |
+          python3 -m pip install openai-whisper torch onnxruntime onnx
+
+      - name: export ${{ matrix.model }}
+        shell: bash
+        run: |
+          cd scripts/whisper
+          python3 ./export-onnx.py --model ${{ matrix.model }}
+          python3 -m onnxruntime.tools.convert_onnx_models_to_ort --optimization_style=Fixed ./
+
+          ls -lh
+
+          ls -lh ~/.cache/whisper
+
+      - name: Publish ${{ matrix.model }} to huggingface
+        shell: bash
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          cd scripts/whisper
+
+          git config --global user.email "csukuangfj@gmail.com"
+          git config --global user.name "Fangjun Kuang"
+
+          GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/csukuangfj/sherpa-onnx-whisper-${{ matrix.model }} huggingface
+
+          cp *.onnx ./huggingface
+          cp *.ort ./huggingface
+          cp *tokens.txt ./huggingface
+
+          cd huggingface
+          git status
+          ls -lh
+          git lfs track "*.onnx"
+          git lfs track "*.ort"
+          git add .
+          git commit -m "upload ${{ matrix.model }}"
+          git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-whisper-${{ matrix.model }} main
--- a/.github/workflows/run-java-test.yaml
+++ b/.github/workflows/run-java-test.yaml
@@ -23,14 +23,14 @@ on:
      - 'sherpa-onnx/jni/*'

 concurrency:
-  group: jni-${{ github.ref }}
+  group: run-java-test-${{ github.ref }}
  cancel-in-progress: true

 permissions:
  contents: read

 jobs:
-  jni:
+  run_java_test:
    runs-on: ${{ matrix.os }}
    strategy:
      fail-fast: false
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.13 FATAL_ERROR)
 project(sherpa-onnx)

-set(SHERPA_ONNX_VERSION "1.5.5")
+set(SHERPA_ONNX_VERSION "1.6.0")

 # Disable warning about
 #
--- a/cmake/kaldi-native-fbank.cmake
+++ b/cmake/kaldi-native-fbank.cmake
@@ -1,9 +1,9 @@
 function(download_kaldi_native_fbank)
  include(FetchContent)

-  set(kaldi_native_fbank_URL  "https://github.com/csukuangfj/kaldi-native-fbank/archive/refs/tags/v1.17.tar.gz")
-  set(kaldi_native_fbank_URL2 "https://huggingface.co/csukuangfj/sherpa-onnx-cmake-deps/resolve/main/kaldi-native-fbank-1.17.tar.gz")
-  set(kaldi_native_fbank_HASH "SHA256=300dc282d51d738e70f194ef13a50bf4cf8d54a3b2686d75f7fc2fb821f8c1e6")
+  set(kaldi_native_fbank_URL  "https://github.com/csukuangfj/kaldi-native-fbank/archive/refs/tags/v1.18.1.tar.gz")
+  set(kaldi_native_fbank_URL2 "https://huggingface.co/csukuangfj/sherpa-onnx-cmake-deps/resolve/main/kaldi-native-fbank-1.18.1.tar.gz")
+  set(kaldi_native_fbank_HASH "SHA256=c7676f319fa97e8c8bca6018792de120895dcfe122fa9b4bff00f8f9165348e7")

  set(KALDI_NATIVE_FBANK_BUILD_TESTS OFF CACHE BOOL "" FORCE)
  set(KALDI_NATIVE_FBANK_BUILD_PYTHON OFF CACHE BOOL "" FORCE)
@@ -12,11 +12,11 @@ function(download_kaldi_native_fbank)
  # If you don't have access to the Internet,
  # please pre-download kaldi-native-fbank
  set(possible_file_locations
-    $ENV{HOME}/Downloads/kaldi-native-fbank-1.17.tar.gz
-    ${PROJECT_SOURCE_DIR}/kaldi-native-fbank-1.17.tar.gz
-    ${PROJECT_BINARY_DIR}/kaldi-native-fbank-1.17.tar.gz
-    /tmp/kaldi-native-fbank-1.17.tar.gz
-    /star-fj/fangjun/download/github/kaldi-native-fbank-1.17.tar.gz
+    $ENV{HOME}/Downloads/kaldi-native-fbank-1.18.1.tar.gz
+    ${PROJECT_SOURCE_DIR}/kaldi-native-fbank-1.18.1.tar.gz
+    ${PROJECT_BINARY_DIR}/kaldi-native-fbank-1.18.1.tar.gz
+    /tmp/kaldi-native-fbank-1.18.1.tar.gz
+    /star-fj/fangjun/download/github/kaldi-native-fbank-1.18.1.tar.gz
  )

  foreach(f IN LISTS possible_file_locations)
--- a/python-api-examples/offline-decode-files.py
+++ b/python-api-examples/offline-decode-files.py
@@ -1,6 +1,7 @@
 #!/usr/bin/env python3
 #
 # Copyright (c)  2023 by manyeyes
+# Copyright (c)  2023  Xiaomi Corporation

 """
 This file demonstrates how to use sherpa-onnx Python API to transcribe
@@ -34,6 +35,27 @@ file(s) with a non-streaming model.

 (3) For CTC models from NeMo

+python3 ./python-api-examples/offline-decode-files.py \
+  --tokens=./sherpa-onnx-nemo-ctc-en-citrinet-512/tokens.txt \
+  --nemo-ctc=./sherpa-onnx-nemo-ctc-en-citrinet-512/model.onnx \
+  --num-threads=2 \
+  --decoding-method=greedy_search \
+  --debug=false \
+  ./sherpa-onnx-nemo-ctc-en-citrinet-512/test_wavs/0.wav \
+  ./sherpa-onnx-nemo-ctc-en-citrinet-512/test_wavs/1.wav \
+  ./sherpa-onnx-nemo-ctc-en-citrinet-512/test_wavs/8k.wav
+
+(4) For Whisper models
+
+python3 ./python-api-examples/offline-decode-files.py \
+  --whisper-encoder=./sherpa-onnx-whisper-base.en/base.en-encoder.int8.onnx \
+  --whisper-decoder=./sherpa-onnx-whisper-base.en/base.en-decoder.int8.onnx \
+  --tokens=./sherpa-onnx-whisper-base.en/base.en-tokens.txt \
+  --num-threads=1 \
+  ./sherpa-onnx-whisper-base.en/test_wavs/0.wav \
+  ./sherpa-onnx-whisper-base.en/test_wavs/1.wav \
+  ./sherpa-onnx-whisper-base.en/test_wavs/8k.wav
+
 Please refer to
 https://k2-fsa.github.io/sherpa/onnx/index.html
 to install sherpa-onnx and to download the pre-trained models
@@ -144,6 +166,20 @@ def get_args():
        help="Number of threads for neural network computation",
    )

+    parser.add_argument(
+        "--whisper-encoder",
+        default="",
+        type=str,
+        help="Path to whisper encoder model",
+    )
+
+    parser.add_argument(
+        "--whisper-decoder",
+        default="",
+        type=str,
+        help="Path to whisper decoder model",
+    )
+
    parser.add_argument(
        "--decoding-method",
        type=str,
@@ -247,6 +283,8 @@ def main():
    if args.encoder:
        assert len(args.paraformer) == 0, args.paraformer
        assert len(args.nemo_ctc) == 0, args.nemo_ctc
+        assert len(args.whisper_encoder) == 0, args.whisper_encoder
+        assert len(args.whisper_decoder) == 0, args.whisper_decoder

        contexts = [x.strip().upper() for x in args.contexts.split("/") if x.strip()]
        if contexts:
@@ -271,6 +309,9 @@ def main():
        )
    elif args.paraformer:
        assert len(args.nemo_ctc) == 0, args.nemo_ctc
+        assert len(args.whisper_encoder) == 0, args.whisper_encoder
+        assert len(args.whisper_decoder) == 0, args.whisper_decoder
+
        assert_file_exists(args.paraformer)

        recognizer = sherpa_onnx.OfflineRecognizer.from_paraformer(
@@ -283,6 +324,11 @@ def main():
            debug=args.debug,
        )
    elif args.nemo_ctc:
+        assert len(args.whisper_encoder) == 0, args.whisper_encoder
+        assert len(args.whisper_decoder) == 0, args.whisper_decoder
+
+        assert_file_exists(args.nemo_ctc)
+
        recognizer = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(
            model=args.nemo_ctc,
            tokens=args.tokens,
@@ -292,6 +338,18 @@ def main():
            decoding_method=args.decoding_method,
            debug=args.debug,
        )
+    elif args.whisper_encoder:
+        assert_file_exists(args.whisper_encoder)
+        assert_file_exists(args.whisper_decoder)
+
+        recognizer = sherpa_onnx.OfflineRecognizer.from_whisper(
+            encoder=args.whisper_encoder,
+            decoder=args.whisper_decoder,
+            tokens=args.tokens,
+            num_threads=args.num_threads,
+            decoding_method=args.decoding_method,
+            debug=args.debug,
+        )
    else:
        print("Please specify at least one model")
        return
--- a/scripts/whisper/.gitignore
+++ b/scripts/whisper/.gitignore
@@ -0,0 +1,4 @@
+*.onnx
+*.config
+*.ort
+*-tokens.txt
--- a/scripts/whisper/README.md
+++ b/scripts/whisper/README.md
@@ -0,0 +1,9 @@
+# Introduction
+
+This folder contains code showing how to convert [Whisper][whisper] to onnx
+and use onnxruntime to replace PyTorch for speech recognition.
+
+You can use [sherpa-onnx][sherpa-onnx] to run the converted model.
+
+[whisper]: https://github.com/openai/whisper
+[sherpa-onnx]: https://github.com/k2-fsa/sherpa-onnx
--- a/scripts/whisper/export-onnx.py
+++ b/scripts/whisper/export-onnx.py
@@ -0,0 +1,439 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+# flake8: noqa
+
+"""
+Note: Code in this file is modified from
+https://github.com/TadaoYamaoka/whisper/blob/main/to_onnx.py
+
+Thanks to https://github.com/TadaoYamaoka
+for making the onnx export script public.
+"""
+
+import argparse
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import onnx
+import torch
+from onnxruntime.quantization import QuantType, quantize_dynamic
+from torch import Tensor, nn
+
+import whisper
+from whisper.model import (
+    AudioEncoder,
+    MultiHeadAttention,
+    ResidualAttentionBlock,
+    TextDecoder,
+)
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        # fmt: off
+        choices=[
+            "tiny", "tiny.en", "base", "base.en",
+            "small", "small.en", "medium", "medium.en",
+            "large", "large-v1", "large-v2"],
+        # fmt: on
+    )
+    return parser.parse_args()
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, Any]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = str(value)
+
+    onnx.save(model, filename)
+
+
+class AudioEncoderTensorCache(nn.Module):
+    def __init__(self, inAudioEncoder: AudioEncoder, inTextDecoder: TextDecoder):
+        super().__init__()
+        self.audioEncoder = inAudioEncoder
+        self.textDecoder = inTextDecoder
+
+    def forward(self, x: Tensor):
+        audio_features = self.audioEncoder(x)
+
+        n_layer_cross_k_list = []
+        n_layer_cross_v_list = []
+        for block in self.textDecoder.blocks:
+            n_layer_cross_k_list.append(block.cross_attn.key(audio_features))
+            n_layer_cross_v_list.append(block.cross_attn.value(audio_features))
+
+        return torch.stack(n_layer_cross_k_list), torch.stack(n_layer_cross_v_list)
+
+
+class MultiHeadAttentionCross(nn.Module):
+    def __init__(self, inMultiHeadAttention: MultiHeadAttention):
+        super().__init__()
+        self.multiHeadAttention = inMultiHeadAttention
+
+    def forward(
+        self,
+        x: Tensor,
+        k: Tensor,
+        v: Tensor,
+        mask: Optional[Tensor] = None,
+    ):
+        q = self.multiHeadAttention.query(x)
+        wv, qk = self.multiHeadAttention.qkv_attention(q, k, v, mask)
+        return self.multiHeadAttention.out(wv)
+
+
+class MultiHeadAttentionSelf(nn.Module):
+    def __init__(self, inMultiHeadAttention: MultiHeadAttention):
+        super().__init__()
+        self.multiHeadAttention = inMultiHeadAttention
+
+    def forward(
+        self,
+        x: Tensor,  # (b, n_ctx      , n_state)
+        k_cache: Tensor,  # (b, n_ctx_cache, n_state)
+        v_cache: Tensor,  # (b, n_ctx_cache, n_state)
+        mask: Tensor,
+    ):
+        q = self.multiHeadAttention.query(x)  # (b, n_ctx, n_state)
+        k = self.multiHeadAttention.key(x)  # (b, n_ctx, n_state)
+        v = self.multiHeadAttention.value(x)  # (b, n_ctx, n_state)
+
+        k_cache[:, -k.shape[1] :, :] = k  # (b, n_ctx_cache + n_ctx, n_state)
+        v_cache[:, -v.shape[1] :, :] = v  # (b, n_ctx_cache + n_ctx, n_state)
+
+        wv, qk = self.multiHeadAttention.qkv_attention(q, k_cache, v_cache, mask)
+        return self.multiHeadAttention.out(wv), k_cache, v_cache
+
+
+class ResidualAttentionBlockTensorCache(nn.Module):
+    def __init__(self, inResidualAttentionBlock: ResidualAttentionBlock):
+        super().__init__()
+        self.originalBlock = inResidualAttentionBlock
+        self.attn = MultiHeadAttentionSelf(inResidualAttentionBlock.attn)
+        self.cross_attn = (
+            MultiHeadAttentionCross(inResidualAttentionBlock.cross_attn)
+            if inResidualAttentionBlock.cross_attn
+            else None
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        self_k_cache: Tensor,
+        self_v_cache: Tensor,
+        cross_k: Tensor,
+        cross_v: Tensor,
+        mask: Tensor,
+    ):
+        self_attn_x, self_k_cache_updated, self_v_cache_updated = self.attn(
+            self.originalBlock.attn_ln(x), self_k_cache, self_v_cache, mask=mask
+        )
+        x = x + self_attn_x
+
+        if self.cross_attn:
+            x = x + self.cross_attn(
+                self.originalBlock.cross_attn_ln(x), cross_k, cross_v
+            )
+
+        x = x + self.originalBlock.mlp(self.originalBlock.mlp_ln(x))
+        return x, self_k_cache_updated, self_v_cache_updated
+
+
+class TextDecoderTensorCache(nn.Module):
+    def __init__(self, inTextDecoder: TextDecoder, in_n_ctx: int):
+        super().__init__()
+        self.textDecoder = inTextDecoder
+        self.n_ctx = in_n_ctx
+
+        self.blocks = []
+        for orginal_block in self.textDecoder.blocks:
+            self.blocks.append(ResidualAttentionBlockTensorCache(orginal_block))
+
+    def forward(
+        self,
+        tokens: Tensor,
+        n_layer_self_k_cache: Tensor,
+        n_layer_self_v_cache: Tensor,
+        n_layer_cross_k: Tensor,
+        n_layer_cross_v: Tensor,
+        offset: Tensor,
+    ):
+        x = (
+            self.textDecoder.token_embedding(tokens)
+            + self.textDecoder.positional_embedding[
+                offset[0] : offset[0] + tokens.shape[-1]
+            ]
+        )
+        x = x.to(n_layer_cross_k[0].dtype)
+
+        i = 0
+        for block in self.blocks:
+            self_k_cache = n_layer_self_k_cache[i, :, : offset[0] + tokens.shape[-1], :]
+            self_v_cache = n_layer_self_v_cache[i, :, : offset[0] + tokens.shape[-1], :]
+            x, self_k_cache, self_v_cache = block(
+                x,
+                self_k_cache=self_k_cache,
+                self_v_cache=self_v_cache,
+                cross_k=n_layer_cross_k[i],
+                cross_v=n_layer_cross_v[i],
+                mask=self.textDecoder.mask,
+            )
+            n_layer_self_k_cache[i, :, : offset[0] + tokens.shape[-1], :] = self_k_cache
+            n_layer_self_v_cache[i, :, : offset[0] + tokens.shape[-1], :] = self_v_cache
+            i += 1
+
+        x = self.textDecoder.ln(x)
+
+        logits = (
+            x
+            @ torch.transpose(self.textDecoder.token_embedding.weight.to(x.dtype), 0, 1)
+        ).float()
+
+        return logits, n_layer_self_k_cache, n_layer_self_v_cache
+
+
+# ref: https://github.com/ggerganov/whisper.cpp/blob/master/models/convert-pt-to-ggml.py#L232
+def convert_tokens(name, model):
+    whisper_dir = Path(whisper.__file__).parent
+    multilingual = model.is_multilingual
+    tokenizer = (
+        whisper_dir
+        / "assets"
+        / (multilingual and "multilingual.tiktoken" or "gpt2.tiktoken")
+    )
+    if not tokenizer.is_file():
+        raise ValueError(f"Cannot find {tokenizer}")
+
+    #  import base64
+
+    with open(tokenizer, "r") as f:
+        contents = f.read()
+        #  tokens = {
+        #      base64.b64decode(token): int(rank)
+        #      for token, rank in (line.split() for line in contents.splitlines() if line)
+        #  }
+        tokens = {
+            token: int(rank)
+            for token, rank in (line.split() for line in contents.splitlines() if line)
+        }
+
+    with open(f"{name}-tokens.txt", "w") as f:
+        for t, i in tokens.items():
+            f.write(f"{t} {i}\n")
+
+
+@torch.no_grad()
+def main():
+    args = get_args()
+    name = args.model
+
+    opset_version = 13
+
+    model = whisper.load_model(name)
+    convert_tokens(name=name, model=model)
+
+    # write tokens
+
+    tokenizer = whisper.tokenizer.get_tokenizer(model.is_multilingual)
+    model.eval()
+    print(model.dims)
+    audio = torch.rand(16000 * 2)
+    audio = whisper.pad_or_trim(audio)
+    assert audio.shape == (16000 * 30,), audio.shape
+
+    # make log-Mel spectrogram and move to the same device as the model
+    mel = whisper.log_mel_spectrogram(audio).to(model.device).unsqueeze(0)
+    batch_size = 1
+    assert mel.shape == (batch_size, 80, 30 * 100)
+
+    encoder = AudioEncoderTensorCache(model.encoder, model.decoder)
+    n_layer_cross_k, n_layer_cross_v = encoder(mel)
+    assert n_layer_cross_k.shape == (
+        model.dims.n_text_layer,
+        batch_size,
+        model.dims.n_audio_ctx,
+        model.dims.n_text_state,
+    ), n_layer_cross_k.shape
+    assert n_layer_cross_v.shape == (
+        model.dims.n_text_layer,
+        batch_size,
+        model.dims.n_audio_ctx,
+        model.dims.n_text_state,
+    ), n_layer_cross_v.shape
+
+    encoder_filename = f"{name}-encoder.onnx"
+    torch.onnx.export(
+        encoder,
+        mel,
+        encoder_filename,
+        opset_version=opset_version,
+        input_names=["mel"],
+        output_names=["n_layer_cross_k", "n_layer_cross_v"],
+        dynamic_axes={
+            "mel": {0: "n_audio"},  # n_audio is also known as batch_size
+            "n_layer_cross_k": {1: "n_audio"},
+            "n_layer_cross_v": {1: "n_audio"},
+        },
+    )
+
+    encoder_meta_data = {
+        "model_type": f"whisper-{name}",
+        "version": "1",
+        "maintainer": "k2-fsa",
+        "n_mels": model.dims.n_mels,
+        "n_audio_ctx": model.dims.n_audio_ctx,
+        "n_audio_state": model.dims.n_audio_state,
+        "n_audio_head": model.dims.n_audio_head,
+        "n_audio_layer": model.dims.n_audio_layer,
+        "n_vocab": model.dims.n_vocab,
+        "n_text_ctx": model.dims.n_text_ctx,
+        "n_text_state": model.dims.n_text_state,
+        "n_text_head": model.dims.n_text_head,
+        "n_text_layer": model.dims.n_text_layer,
+        "sot_sequence": ",".join(list(map(str, tokenizer.sot_sequence))),
+        "all_language_tokens": ",".join(list(map(str, tokenizer.all_language_tokens))),
+        "all_language_codes": ",".join(tokenizer.all_language_codes),
+        "sot": tokenizer.sot,
+        "sot_index": tokenizer.sot_sequence.index(tokenizer.sot),
+        "eot": tokenizer.eot,
+        "blank_id": tokenizer.encode(" ")[0],
+        "is_multilingual": int(model.is_multilingual),
+        "no_speech": tokenizer.no_speech,
+        "non_speech_tokens": ",".join(list(map(str, tokenizer.non_speech_tokens))),
+        "transcribe": tokenizer.transcribe,
+        "translate": tokenizer.translate,
+        "sot_prev": tokenizer.sot_prev,
+        "sot_lm": tokenizer.sot_lm,
+        "no_timestamps": tokenizer.no_timestamps,
+    }
+    print(f"encoder_meta_data: {encoder_meta_data}")
+    add_meta_data(filename=encoder_filename, meta_data=encoder_meta_data)
+
+    n_audio = mel.shape[0]
+    tokens = torch.tensor([[tokenizer.sot, tokenizer.sot, tokenizer.sot]] * n_audio).to(
+        mel.device
+    )  # [n_audio, 3]
+    decoder = TextDecoderTensorCache(model.decoder, model.dims.n_text_ctx)
+    n_layer_self_k_cache = torch.zeros(
+        (
+            len(model.decoder.blocks),
+            n_audio,
+            model.dims.n_text_ctx,
+            model.dims.n_text_state,
+        ),
+        device=mel.device,
+    )
+    n_layer_self_v_cache = torch.zeros(
+        (
+            len(model.decoder.blocks),
+            n_audio,
+            model.dims.n_text_ctx,
+            model.dims.n_text_state,
+        ),
+        device=mel.device,
+    )
+    offset = torch.zeros(1, dtype=torch.int64).to(mel.device)
+    logits, n_layer_self_k_cache, n_layer_self_v_cache = decoder(
+        tokens,
+        n_layer_self_k_cache,
+        n_layer_self_v_cache,
+        n_layer_cross_k,
+        n_layer_cross_v,
+        offset,
+    )
+    assert logits.shape == (n_audio, tokens.shape[1], model.dims.n_vocab)
+    assert n_layer_self_k_cache.shape == (
+        model.dims.n_text_layer,
+        n_audio,
+        model.dims.n_text_ctx,
+        model.dims.n_text_state,
+    )
+    assert n_layer_self_v_cache.shape == (
+        model.dims.n_text_layer,
+        n_audio,
+        model.dims.n_text_ctx,
+        model.dims.n_text_state,
+    )
+
+    offset = torch.tensor([tokens.shape[1]], dtype=torch.int64).to(mel.device)
+    tokens = torch.tensor([[tokenizer.sot]] * n_audio).to(mel.device)  # [n_audio, 1]
+
+    logits, out_n_layer_self_k_cache, out_n_layer_self_v_cache = decoder(
+        tokens,
+        n_layer_self_k_cache,
+        n_layer_self_v_cache,
+        n_layer_cross_k,
+        n_layer_cross_v,
+        offset,
+    )
+
+    decoder_filename = f"{name}-decoder.onnx"
+    torch.onnx.export(
+        decoder,
+        (
+            tokens,
+            n_layer_self_k_cache,
+            n_layer_self_v_cache,
+            n_layer_cross_k,
+            n_layer_cross_v,
+            offset,
+        ),
+        decoder_filename,
+        opset_version=opset_version,
+        input_names=[
+            "tokens",
+            "in_n_layer_self_k_cache",
+            "in_n_layer_self_v_cache",
+            "n_layer_cross_k",
+            "n_layer_cross_v",
+            "offset",
+        ],
+        output_names=["logits", "out_n_layer_self_k_cache", "out_n_layer_self_v_cache"],
+        dynamic_axes={
+            "tokens": {0: "n_audio", 1: "n_tokens"},
+            "in_n_layer_self_k_cache": {1: "n_audio"},
+            "in_n_layer_self_v_cache": {1: "n_audio"},
+            "n_layer_cross_k": {1: "n_audio"},
+            "n_layer_cross_v": {1: "n_audio"},
+        },
+    )
+
+    # Generate int8 quantization models
+    # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection
+
+    print("Generate int8 quantization models")
+
+    encoder_filename_int8 = f"{name}-encoder.int8.onnx"
+    quantize_dynamic(
+        model_input=encoder_filename,
+        model_output=encoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+    decoder_filename_int8 = f"{name}-decoder.int8.onnx"
+    quantize_dynamic(
+        model_input=decoder_filename,
+        model_output=decoder_filename_int8,
+        op_types_to_quantize=["MatMul"],
+        weight_type=QuantType.QInt8,
+    )
+
+
+if __name__ == "__main__":
+    main()
--- a/scripts/whisper/requirements.txt
+++ b/scripts/whisper/requirements.txt
@@ -0,0 +1 @@
+openai-whisper
--- a/scripts/whisper/test.py
+++ b/scripts/whisper/test.py
@@ -0,0 +1,241 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Fangjun Kuang)
+"""
+Please first run ./export-onnx.py
+before you run this script
+"""
+import base64
+from typing import Tuple
+
+import kaldi_native_fbank as knf
+import onnxruntime as ort
+import torch
+
+import whisper
+import argparse
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        # fmt: off
+        choices=[
+            "tiny", "tiny.en", "base", "base.en",
+            "small", "small.en", "medium", "medium.en",
+            "large", "large-v1", "large-v2"],
+        # fmt: on
+    )
+    return parser.parse_args()
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        encoder: str,
+        decoder: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 4
+
+        self.session_opts = session_opts
+
+        self.init_encoder(encoder)
+        self.init_decoder(decoder)
+
+    def init_encoder(self, encoder: str):
+        self.encoder = ort.InferenceSession(
+            encoder,
+            sess_options=self.session_opts,
+        )
+
+        meta = self.encoder.get_modelmeta().custom_metadata_map
+        self.n_text_layer = int(meta["n_text_layer"])
+        self.n_text_ctx = int(meta["n_text_ctx"])
+        self.n_text_state = int(meta["n_text_state"])
+        self.sot = int(meta["sot"])
+        self.eot = int(meta["eot"])
+        self.translate = int(meta["translate"])
+        self.no_timestamps = int(meta["no_timestamps"])
+        self.no_speech = int(meta["no_speech"])
+        self.blank = int(meta["blank_id"])
+
+        self.sot_sequence = list(map(int, meta["sot_sequence"].split(",")))
+
+        self.is_multilingual = int(meta["is_multilingual"]) == 1
+
+    def init_decoder(self, decoder: str):
+        self.decoder = ort.InferenceSession(
+            decoder,
+            sess_options=self.session_opts,
+        )
+
+    def run_encoder(
+        self,
+        mel: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        n_layer_cross_k, n_layer_cross_v = self.encoder.run(
+            [
+                self.encoder.get_outputs()[0].name,
+                self.encoder.get_outputs()[1].name,
+            ],
+            {
+                self.encoder.get_inputs()[0].name: mel.numpy(),
+            },
+        )
+        return torch.from_numpy(n_layer_cross_k), torch.from_numpy(n_layer_cross_v)
+
+    def run_decoder(
+        self,
+        tokens: torch.Tensor,
+        n_layer_self_k_cache: torch.Tensor,
+        n_layer_self_v_cache: torch.Tensor,
+        n_layer_cross_k: torch.Tensor,
+        n_layer_cross_v: torch.Tensor,
+        offset: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        logits, out_n_layer_self_k_cache, out_n_layer_self_v_cache = self.decoder.run(
+            [
+                self.decoder.get_outputs()[0].name,
+                self.decoder.get_outputs()[1].name,
+                self.decoder.get_outputs()[2].name,
+            ],
+            {
+                self.decoder.get_inputs()[0].name: tokens.numpy(),
+                self.decoder.get_inputs()[1].name: n_layer_self_k_cache.numpy(),
+                self.decoder.get_inputs()[2].name: n_layer_self_v_cache.numpy(),
+                self.decoder.get_inputs()[3].name: n_layer_cross_k.numpy(),
+                self.decoder.get_inputs()[4].name: n_layer_cross_v.numpy(),
+                self.decoder.get_inputs()[5].name: offset.numpy(),
+            },
+        )
+        return (
+            torch.from_numpy(logits),
+            torch.from_numpy(out_n_layer_self_k_cache),
+            torch.from_numpy(out_n_layer_self_v_cache),
+        )
+
+    def get_self_cache(self) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size = 1
+        n_layer_self_k_cache = torch.zeros(
+            self.n_text_layer,
+            batch_size,
+            self.n_text_ctx,
+            self.n_text_state,
+        )
+        n_layer_self_v_cache = torch.zeros(
+            self.n_text_layer,
+            batch_size,
+            self.n_text_ctx,
+            self.n_text_state,
+        )
+        return n_layer_self_k_cache, n_layer_self_v_cache
+
+    def suppress_tokens(self, logits, is_initial: bool) -> None:
+        # suppress blank
+        if is_initial:
+            logits[self.eot] = float("-inf")
+            logits[self.blank] = float("-inf")
+
+        # suppress <|notimestamps|>
+        logits[self.no_timestamps] = float("-inf")
+
+        logits[self.sot] = float("-inf")
+        logits[self.no_speech] = float("-inf")
+
+        # logits is changed in-place
+        logits[self.translate] = float("-inf")
+
+
+def load_tokens(filename):
+    tokens = dict()
+    with open(filename, "r") as f:
+        for line in f:
+            t, i = line.split()
+            tokens[int(i)] = t
+    return tokens
+
+
+def main():
+    args = get_args()
+    name = args.model
+
+    encoder = f"./{name}-encoder.onnx"
+    decoder = f"./{name}-decoder.onnx"
+    audio = whisper.load_audio("0.wav")
+
+    features = []
+    online_whisper_fbank = knf.OnlineWhisperFbank(knf.FrameExtractionOptions())
+    online_whisper_fbank.accept_waveform(16000, audio)
+    online_whisper_fbank.input_finished()
+    for i in range(online_whisper_fbank.num_frames_ready):
+        f = online_whisper_fbank.get_frame(i)
+        f = torch.from_numpy(f)
+        features.append(f)
+
+    features = torch.stack(features)
+
+    log_spec = torch.clamp(features, min=1e-10).log10()
+    log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+    mel = (log_spec + 4.0) / 4.0
+    target = 3000
+    mel = torch.nn.functional.pad(mel, (0, 0, 0, target - mel.shape[0]), "constant", 0)
+    mel = mel.t().unsqueeze(0)
+
+    model = OnnxModel(encoder, decoder)
+    n_layer_cross_k, n_layer_cross_v = model.run_encoder(mel)
+    n_layer_self_k_cache, n_layer_self_v_cache = model.get_self_cache()
+
+    tokens = torch.tensor([model.sot_sequence], dtype=torch.int64)
+    offset = torch.zeros(1, dtype=torch.int64)
+    logits, n_layer_self_k_cache, n_layer_self_v_cache = model.run_decoder(
+        tokens=tokens,
+        n_layer_self_k_cache=n_layer_self_k_cache,
+        n_layer_self_v_cache=n_layer_self_v_cache,
+        n_layer_cross_k=n_layer_cross_k,
+        n_layer_cross_v=n_layer_cross_v,
+        offset=offset,
+    )
+    # logits.shape (batch_size, tokens.shape[1], vocab_size)
+    logits = logits[0, -1]
+    model.suppress_tokens(logits, is_initial=True)
+    #  logits = logits.softmax(dim=-1)
+    # for greedy search, we don't need to compute softmax or log_softmax
+    max_token_id = logits.argmax(dim=-1)
+    results = []
+    for i in range(model.n_text_ctx):
+        if max_token_id == model.eot:
+            break
+        results.append(max_token_id.item())
+        tokens = torch.tensor([[results[-1]]])
+        offset += 1
+
+        logits, n_layer_self_k_cache, n_layer_self_v_cache = model.run_decoder(
+            tokens=tokens,
+            n_layer_self_k_cache=n_layer_self_k_cache,
+            n_layer_self_v_cache=n_layer_self_v_cache,
+            n_layer_cross_k=n_layer_cross_k,
+            n_layer_cross_v=n_layer_cross_v,
+            offset=offset,
+        )
+        logits = logits[0, -1]
+        model.suppress_tokens(logits, is_initial=False)
+        max_token_id = logits.argmax(dim=-1)
+    token_table = load_tokens(f"./{name}-tokens.txt")
+    s = b""
+    for i in results:
+        if i in token_table:
+            s += base64.b64decode(token_table[i])
+        else:
+            print("oov", i)
+
+    print(s.decode().strip())
+    print(results)
+    print(model.sot_sequence)
+
+
+if __name__ == "__main__":
+    main()
--- a/sherpa-onnx/csrc/CMakeLists.txt
+++ b/sherpa-onnx/csrc/CMakeLists.txt
@@ -11,6 +11,7 @@ if(SHERPA_ONNX_ENABLE_PYTHON)
 endif()

 set(sources
+  base64-decode.cc
  cat.cc
  context-graph.cc
  endpoint.cc
@@ -35,6 +36,9 @@ set(sources
  offline-transducer-model-config.cc
  offline-transducer-model.cc
  offline-transducer-modified-beam-search-decoder.cc
+  offline-whisper-greedy-search-decoder.cc
+  offline-whisper-model-config.cc
+  offline-whisper-model.cc
  online-conformer-transducer-model.cc
  online-lm-config.cc
  online-lm.cc
@@ -50,12 +54,12 @@ set(sources
  online-zipformer-transducer-model.cc
  online-zipformer2-transducer-model.cc
  onnx-utils.cc
-  session.cc
  packed-sequence.cc
  pad-sequence.cc
  parse-options.cc
  provider.cc
  resample.cc
+  session.cc
  slice.cc
  stack.cc
  symbol-table.cc
--- a/sherpa-onnx/csrc/base64-decode.cc
+++ b/sherpa-onnx/csrc/base64-decode.cc
@@ -0,0 +1,67 @@
+// sherpa-onnx/csrc/base64-decode.cc
+//
+// Copyright (c)  2022-2023  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/base64-decode.h"
+
+#include "sherpa-onnx/csrc/macros.h"
+
+namespace sherpa_onnx {
+
+static int32_t Ord(char c) {
+  if (c >= 'A' && c <= 'Z') {
+    return c - 'A';
+  } else if (c >= 'a' && c <= 'z') {
+    return c - 'a' + ('Z' - 'A') + 1;
+  } else if (c >= '0' && c <= '9') {
+    return c - '0' + ('Z' - 'A') + ('z' - 'a') + 2;
+  } else if (c == '+') {
+    return 62;
+  } else if (c == '/') {
+    return 63;
+  }
+
+  SHERPA_ONNX_LOGE("Unknown character %d, %c\n", c, c);
+
+  exit(-1);
+}
+
+// see
+// https://github.com/ReneNyffenegger/cpp-base64/blob/master/base64.cpp#L243
+std::string Base64Decode(const std::string &s) {
+  if (s.empty()) {
+    SHERPA_ONNX_LOGE("Empty string!");
+    exit(-1);
+  }
+
+  int32_t n = s.size() / 4 * 3;
+
+  std::string ans;
+  ans.reserve(n);
+
+  int32_t i = 0;
+  while (i < static_cast<int32_t>(s.size())) {
+    if (s[i] == '=') {
+      return " ";
+    }
+
+    int32_t first = (Ord(s[i]) << 2) + ((Ord(s[i + 1]) & 0x30) >> 4);
+    ans.push_back(first);
+
+    if (i + 2 < static_cast<int32_t>(s.size()) && s[i + 2] != '=') {
+      int32_t second =
+          ((Ord(s[i + 1]) & 0x0f) << 4) + ((Ord(s[i + 2]) & 0x3c) >> 2);
+      ans.push_back(second);
+
+      if (i + 3 < static_cast<int32_t>(s.size()) && s[i + 3] != '=') {
+        int32_t third = ((Ord(s[i + 2]) & 0x03) << 6) + Ord(s[i + 3]);
+        ans.push_back(third);
+      }
+    }
+    i += 4;
+  }
+
+  return ans;
+}
+
+}  // namespace sherpa_onnx
--- a/sherpa-onnx/csrc/base64-decode.h
+++ b/sherpa-onnx/csrc/base64-decode.h
@@ -0,0 +1,19 @@
+// sherpa-onnx/csrc/base64-decode.h
+//
+// Copyright (c)  2022-2023  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_CSRC_BASE64_DECODE_H_
+#define SHERPA_ONNX_CSRC_BASE64_DECODE_H_
+
+#include <string>
+
+namespace sherpa_onnx {
+
+/** @param s A base64 encoded string.
+ *  @return Return the decoded string.
+ */
+std::string Base64Decode(const std::string &s);
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_BASE64_DECODE_H_
--- a/sherpa-onnx/csrc/macros.h
+++ b/sherpa-onnx/csrc/macros.h
@@ -1,4 +1,3 @@
-
 // sherpa-onnx/csrc/macros.h
 //
 // Copyright      2023  Xiaomi Corporation
--- a/sherpa-onnx/csrc/offline-model-config.cc
+++ b/sherpa-onnx/csrc/offline-model-config.cc
@@ -14,6 +14,7 @@ void OfflineModelConfig::Register(ParseOptions *po) {
  transducer.Register(po);
  paraformer.Register(po);
  nemo_ctc.Register(po);
+  whisper.Register(po);

  po->Register("tokens", &tokens, "Path to tokens.txt");

@@ -28,7 +29,7 @@ void OfflineModelConfig::Register(ParseOptions *po) {

  po->Register("model-type", &model_type,
               "Specify it to reduce model initialization time. "
-               "Valid values are: transducer, paraformer, nemo_ctc. "
+               "Valid values are: transducer, paraformer, nemo_ctc, whisper."
               "All other values lead to loading the model twice.");
 }

@@ -51,6 +52,10 @@ bool OfflineModelConfig::Validate() const {
    return nemo_ctc.Validate();
  }

+  if (!whisper.encoder.empty()) {
+    return whisper.Validate();
+  }
+
  return transducer.Validate();
 }

@@ -61,6 +66,7 @@ std::string OfflineModelConfig::ToString() const {
  os << "transducer=" << transducer.ToString() << ", ";
  os << "paraformer=" << paraformer.ToString() << ", ";
  os << "nemo_ctc=" << nemo_ctc.ToString() << ", ";
+  os << "whisper=" << whisper.ToString() << ", ";
  os << "tokens=\"" << tokens << "\", ";
  os << "num_threads=" << num_threads << ", ";
  os << "debug=" << (debug ? "True" : "False") << ", ";
--- a/sherpa-onnx/csrc/offline-model-config.h
+++ b/sherpa-onnx/csrc/offline-model-config.h
@@ -9,6 +9,7 @@
 #include "sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.h"
 #include "sherpa-onnx/csrc/offline-paraformer-model-config.h"
 #include "sherpa-onnx/csrc/offline-transducer-model-config.h"
+#include "sherpa-onnx/csrc/offline-whisper-model-config.h"

 namespace sherpa_onnx {

@@ -16,6 +17,7 @@ struct OfflineModelConfig {
  OfflineTransducerModelConfig transducer;
  OfflineParaformerModelConfig paraformer;
  OfflineNemoEncDecCtcModelConfig nemo_ctc;
+  OfflineWhisperModelConfig whisper;

  std::string tokens;
  int32_t num_threads = 2;
@@ -37,11 +39,13 @@ struct OfflineModelConfig {
  OfflineModelConfig(const OfflineTransducerModelConfig &transducer,
                     const OfflineParaformerModelConfig &paraformer,
                     const OfflineNemoEncDecCtcModelConfig &nemo_ctc,
+                     const OfflineWhisperModelConfig &whisper,
                     const std::string &tokens, int32_t num_threads, bool debug,
                     const std::string &provider, const std::string &model_type)
      : transducer(transducer),
        paraformer(paraformer),
        nemo_ctc(nemo_ctc),
+        whisper(whisper),
        tokens(tokens),
        num_threads(num_threads),
        debug(debug),
--- a/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.cc
+++ b/sherpa-onnx/csrc/offline-nemo-enc-dec-ctc-model-config.cc
@@ -16,7 +16,7 @@ void OfflineNemoEncDecCtcModelConfig::Register(ParseOptions *po) {

 bool OfflineNemoEncDecCtcModelConfig::Validate() const {
  if (!FileExists(model)) {
-    SHERPA_ONNX_LOGE("%s does not exist", model.c_str());
+    SHERPA_ONNX_LOGE("NeMo model: %s does not exist", model.c_str());
    return false;
  }

--- a/sherpa-onnx/csrc/offline-paraformer-model-config.cc
+++ b/sherpa-onnx/csrc/offline-paraformer-model-config.cc
@@ -15,7 +15,7 @@ void OfflineParaformerModelConfig::Register(ParseOptions *po) {

 bool OfflineParaformerModelConfig::Validate() const {
  if (!FileExists(model)) {
-    SHERPA_ONNX_LOGE("%s does not exist", model.c_str());
+    SHERPA_ONNX_LOGE("Paraformer model %s does not exist", model.c_str());
    return false;
  }

--- a/sherpa-onnx/csrc/offline-recognizer-impl.cc
+++ b/sherpa-onnx/csrc/offline-recognizer-impl.cc
@@ -11,6 +11,7 @@
 #include "sherpa-onnx/csrc/offline-recognizer-ctc-impl.h"
 #include "sherpa-onnx/csrc/offline-recognizer-paraformer-impl.h"
 #include "sherpa-onnx/csrc/offline-recognizer-transducer-impl.h"
+#include "sherpa-onnx/csrc/offline-recognizer-whisper-impl.h"
 #include "sherpa-onnx/csrc/onnx-utils.h"
 #include "sherpa-onnx/csrc/text-utils.h"

@@ -26,6 +27,8 @@ std::unique_ptr<OfflineRecognizerImpl> OfflineRecognizerImpl::Create(
      return std::make_unique<OfflineRecognizerParaformerImpl>(config);
    } else if (model_type == "nemo_ctc") {
      return std::make_unique<OfflineRecognizerCtcImpl>(config);
+    } else if (model_type == "whisper") {
+      return std::make_unique<OfflineRecognizerWhisperImpl>(config);
    } else {
      SHERPA_ONNX_LOGE(
          "Invalid model_type: %s. Trying to load the model to get its type",
@@ -43,6 +46,8 @@ std::unique_ptr<OfflineRecognizerImpl> OfflineRecognizerImpl::Create(
    model_filename = config.model_config.paraformer.model;
  } else if (!config.model_config.nemo_ctc.model.empty()) {
    model_filename = config.model_config.nemo_ctc.model;
+  } else if (!config.model_config.whisper.encoder.empty()) {
+    model_filename = config.model_config.whisper.encoder;
  } else {
    SHERPA_ONNX_LOGE("Please provide a model");
    exit(-1);
@@ -77,6 +82,8 @@ std::unique_ptr<OfflineRecognizerImpl> OfflineRecognizerImpl::Create(
        "\n    "
        "https://huggingface.co/csukuangfj/"
        "paraformer-onnxruntime-python-example/blob/main/add-model-metadata.py"
+        "\n    "
+        "(3) Whisper"
        "\n");
    exit(-1);
  }
@@ -95,12 +102,17 @@ std::unique_ptr<OfflineRecognizerImpl> OfflineRecognizerImpl::Create(
    return std::make_unique<OfflineRecognizerCtcImpl>(config);
  }

+  if (strncmp(model_type.c_str(), "whisper", 7) == 0) {
+    return std::make_unique<OfflineRecognizerWhisperImpl>(config);
+  }
+
  SHERPA_ONNX_LOGE(
      "\nUnsupported model_type: %s\n"
      "We support only the following model types at present: \n"
      " - Non-streaming transducer models from icefall\n"
      " - Non-streaming Paraformer models from FunASR\n"
-      " - EncDecCTCModelBPE models from NeMo\n",
+      " - EncDecCTCModelBPE models from NeMo\n"
+      " - Whisper models\n",
      model_type.c_str());

  exit(-1);
--- a/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
+++ b/sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
@@ -0,0 +1,152 @@
+// sherpa-onnx/csrc/offline-recognizer-whisper-impl.h
+//
+// Copyright (c)  2022-2023  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
+#define SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
+
+#include <algorithm>
+#include <cmath>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "sherpa-onnx/csrc/offline-model-config.h"
+#include "sherpa-onnx/csrc/offline-recognizer-impl.h"
+#include "sherpa-onnx/csrc/offline-recognizer.h"
+#include "sherpa-onnx/csrc/offline-whisper-decoder.h"
+#include "sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h"
+#include "sherpa-onnx/csrc/offline-whisper-model.h"
+#include "sherpa-onnx/csrc/symbol-table.h"
+#include "sherpa-onnx/csrc/transpose.h"
+
+namespace sherpa_onnx {
+
+static OfflineRecognitionResult Convert(const OfflineWhisperDecoderResult &src,
+                                        const SymbolTable &sym_table) {
+  OfflineRecognitionResult r;
+  r.tokens.reserve(src.tokens.size());
+
+  for (auto i : src.tokens) {
+    if (!sym_table.contains(i)) {
+      continue;
+    }
+
+    const auto &s = sym_table[i];
+    r.text += s;
+    r.tokens.push_back(s);
+  }
+
+  return r;
+}
+
+class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
+ public:
+  explicit OfflineRecognizerWhisperImpl(const OfflineRecognizerConfig &config)
+      : config_(config),
+        symbol_table_(config_.model_config.tokens),
+        model_(std::make_unique<OfflineWhisperModel>(config.model_config)) {
+    // tokens.txt from whisper is base64 encoded, so we need to decode it
+    symbol_table_.ApplyBase64Decode();
+
+    if (config.decoding_method == "greedy_search") {
+      decoder_ =
+          std::make_unique<OfflineWhisperGreedySearchDecoder>(model_.get());
+    } else {
+      SHERPA_ONNX_LOGE(
+          "Only greedy_search is supported at present for whisper. Given %s",
+          config.decoding_method.c_str());
+      exit(-1);
+    }
+  }
+
+  std::unique_ptr<OfflineStream> CreateStream() const override {
+    return std::make_unique<OfflineStream>(WhisperTag{});
+  }
+
+  void DecodeStreams(OfflineStream **ss, int32_t n) const override {
+    // batch decoding is not implemented yet
+    for (int32_t i = 0; i != n; ++i) {
+      DecodeStream(ss[i]);
+    }
+  }
+
+ private:
+  void DecodeStream(OfflineStream *s) const {
+    int32_t max_num_frames = 3000;
+    auto memory_info =
+        Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
+
+    int32_t feat_dim = s->FeatureDim();
+    std::vector<float> f = s->GetFrames();
+    int32_t num_frames = f.size() / feat_dim;
+
+    if (num_frames > max_num_frames) {
+      SHERPA_ONNX_LOGE("Only waves less than 30 seconds are supported.");
+      exit(-1);
+    }
+
+    NormalizeFeatures(f.data(), num_frames, feat_dim);
+
+    std::array<int64_t, 3> shape{1, max_num_frames, feat_dim};
+
+    Ort::Value mel = Ort::Value::CreateTensor<float>(
+        model_->Allocator(), shape.data(), shape.size());
+    float *p_mel = mel.GetTensorMutableData<float>();
+    std::copy(f.begin(), f.end(), p_mel);
+
+    memset(p_mel + f.size(), 0,
+           (max_num_frames - num_frames) * feat_dim * sizeof(float));
+    mel = Transpose12(model_->Allocator(), &mel);
+
+    auto cross_kv = model_->ForwardEncoder(std::move(mel));
+    auto results =
+        decoder_->Decode(std::move(cross_kv.first), std::move(cross_kv.second));
+
+    auto r = Convert(results[0], symbol_table_);
+    s->SetResult(r);
+  }
+
+ private:
+  static void NormalizeFeatures(float *features, int32_t num_frames,
+                                int32_t feat_dim) {
+    // log_spec = torch.clamp(features, min=1e-10).log10()
+    // log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
+    // mel = (log_spec + 4.0) / 4.0
+
+    int32_t n = num_frames * feat_dim;
+    float max_v = -1e20;
+    for (int32_t i = 0; i != n; ++i) {
+      float f = features[i];
+
+      f = std::max<float>(f, 1e-10);
+      f = std::log10(f);
+
+      max_v = std::max(f, max_v);
+
+      features[i] = f;
+    }
+
+    max_v -= 8;
+
+    for (int32_t i = 0; i != n; ++i) {
+      float f = features[i];
+      f = std::max(f, max_v);
+
+      f = (f + 4) / 4;
+
+      features[i] = f;
+    }
+  }
+
+ private:
+  OfflineRecognizerConfig config_;
+  SymbolTable symbol_table_;
+  std::unique_ptr<OfflineWhisperModel> model_;
+  std::unique_ptr<OfflineWhisperDecoder> decoder_;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_OFFLINE_RECOGNIZER_WHISPER_IMPL_H_
--- a/sherpa-onnx/csrc/offline-stream.cc
+++ b/sherpa-onnx/csrc/offline-stream.cc
@@ -86,6 +86,15 @@ class OfflineStream::Impl {
    fbank_ = std::make_unique<knf::OnlineFbank>(opts_);
  }

+  Impl(WhisperTag /*tag*/, ContextGraphPtr context_graph)
+      : context_graph_(context_graph) {
+    config_.normalize_samples = true;
+    opts_.frame_opts.samp_freq = 16000;
+    opts_.mel_opts.num_bins = 80;
+    whisper_fbank_ =
+        std::make_unique<knf::OnlineWhisperFbank>(opts_.frame_opts);
+  }
+
  void AcceptWaveform(int32_t sampling_rate, const float *waveform, int32_t n) {
    if (config_.normalize_samples) {
      AcceptWaveformImpl(sampling_rate, waveform, n);
@@ -117,20 +126,35 @@ class OfflineStream::Impl {
          lowpass_filter_width);
      std::vector<float> samples;
      resampler->Resample(waveform, n, true, &samples);
+
+      if (fbank_) {
        fbank_->AcceptWaveform(opts_.frame_opts.samp_freq, samples.data(),
                               samples.size());
        fbank_->InputFinished();
-      return;
+      } else {
+        whisper_fbank_->AcceptWaveform(opts_.frame_opts.samp_freq,
+                                       samples.data(), samples.size());
+        whisper_fbank_->InputFinished();
      }

+      return;
+    }  // if (sampling_rate != opts_.frame_opts.samp_freq)
+
+    if (fbank_) {
      fbank_->AcceptWaveform(sampling_rate, waveform, n);
      fbank_->InputFinished();
+    } else {
+      whisper_fbank_->AcceptWaveform(sampling_rate, waveform, n);
+      whisper_fbank_->InputFinished();
+    }
  }

  int32_t FeatureDim() const { return opts_.mel_opts.num_bins; }

  std::vector<float> GetFrames() const {
-    int32_t n = fbank_->NumFramesReady();
+    int32_t n =
+        fbank_ ? fbank_->NumFramesReady() : whisper_fbank_->NumFramesReady();
+
    assert(n > 0 && "Please first call AcceptWaveform()");

    int32_t feature_dim = FeatureDim();
@@ -140,7 +164,8 @@ class OfflineStream::Impl {
    float *p = features.data();

    for (int32_t i = 0; i != n; ++i) {
-      const float *f = fbank_->GetFrame(i);
+      const float *f =
+          fbank_ ? fbank_->GetFrame(i) : whisper_fbank_->GetFrame(i);
      std::copy(f, f + feature_dim, p);
      p += feature_dim;
    }
@@ -191,6 +216,7 @@ class OfflineStream::Impl {
 private:
  OfflineFeatureExtractorConfig config_;
  std::unique_ptr<knf::OnlineFbank> fbank_;
+  std::unique_ptr<knf::OnlineWhisperFbank> whisper_fbank_;
  knf::FbankOptions opts_;
  OfflineRecognitionResult r_;
  ContextGraphPtr context_graph_;
@@ -201,6 +227,10 @@ OfflineStream::OfflineStream(
    ContextGraphPtr context_graph /*= nullptr*/)
    : impl_(std::make_unique<Impl>(config, context_graph)) {}

+OfflineStream::OfflineStream(WhisperTag tag,
+                             ContextGraphPtr context_graph /*= nullptr*/)
+    : impl_(std::make_unique<Impl>(tag, context_graph)) {}
+
 OfflineStream::~OfflineStream() = default;

 void OfflineStream::AcceptWaveform(int32_t sampling_rate, const float *waveform,
--- a/sherpa-onnx/csrc/offline-stream.h
+++ b/sherpa-onnx/csrc/offline-stream.h
@@ -65,10 +65,15 @@ struct OfflineFeatureExtractorConfig {
  void Register(ParseOptions *po);
 };

+struct WhisperTag {};
+
 class OfflineStream {
 public:
  explicit OfflineStream(const OfflineFeatureExtractorConfig &config = {},
                         ContextGraphPtr context_graph = nullptr);
+
+  explicit OfflineStream(WhisperTag tag,
+                         ContextGraphPtr context_graph = nullptr);
  ~OfflineStream();

  /**
--- a/sherpa-onnx/csrc/offline-transducer-model-config.cc
+++ b/sherpa-onnx/csrc/offline-transducer-model-config.cc
@@ -18,17 +18,20 @@ void OfflineTransducerModelConfig::Register(ParseOptions *po) {

 bool OfflineTransducerModelConfig::Validate() const {
  if (!FileExists(encoder_filename)) {
-    SHERPA_ONNX_LOGE("encoder: %s does not exist", encoder_filename.c_str());
+    SHERPA_ONNX_LOGE("transducer encoder: %s does not exist",
+                     encoder_filename.c_str());
    return false;
  }

  if (!FileExists(decoder_filename)) {
-    SHERPA_ONNX_LOGE("decoder: %s does not exist", decoder_filename.c_str());
+    SHERPA_ONNX_LOGE("transducer decoder: %s does not exist",
+                     decoder_filename.c_str());
    return false;
  }

  if (!FileExists(joiner_filename)) {
-    SHERPA_ONNX_LOGE("joiner: %s does not exist", joiner_filename.c_str());
+    SHERPA_ONNX_LOGE("transducer joiner: %s does not exist",
+                     joiner_filename.c_str());
    return false;
  }

--- a/sherpa-onnx/csrc/offline-whisper-decoder.h
+++ b/sherpa-onnx/csrc/offline-whisper-decoder.h
@@ -0,0 +1,38 @@
+// sherpa-onnx/csrc/offline-whisper-decoder.h
+//
+// Copyright (c)  2023  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_DECODER_H_
+#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_DECODER_H_
+
+#include <vector>
+
+#include "onnxruntime_cxx_api.h"  // NOLINT
+
+namespace sherpa_onnx {
+
+struct OfflineWhisperDecoderResult {
+  /// The decoded token IDs
+  std::vector<int32_t> tokens;
+};
+
+class OfflineWhisperDecoder {
+ public:
+  virtual ~OfflineWhisperDecoder() = default;
+
+  /** Run beam search given the output from the whisper encoder model.
+   *
+   * @param n_layer_cross_k       A 4-D tensor of shape
+   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
+   * @param n_layer_cross_v       A 4-D tensor of shape
+   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
+   *
+   * @return Return a vector of size `N` containing the decoded results.
+   */
+  virtual std::vector<OfflineWhisperDecoderResult> Decode(
+      Ort::Value n_layer_cross_k, Ort::Value n_layer_cross_v) = 0;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_DECODER_H_
--- a/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.cc
+++ b/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.cc
@@ -0,0 +1,93 @@
+// sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.cc
+//
+// Copyright (c)  2023  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h"
+
+#include <algorithm>
+#include <utility>
+
+namespace sherpa_onnx {
+
+std::vector<OfflineWhisperDecoderResult>
+OfflineWhisperGreedySearchDecoder::Decode(Ort::Value cross_k,
+                                          Ort::Value cross_v) {
+  auto memory_info =
+      Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
+
+  auto self_kv_cache = model_->GetInitialSelfKVCache();
+
+  std::vector<int64_t> initial_tokens = model_->GetInitialTokens();
+  int32_t batch_size = 1;
+  std::array<int64_t, 2> token_shape{
+      batch_size, static_cast<int64_t>(initial_tokens.size())};
+
+  Ort::Value tokens = Ort::Value::CreateTensor(
+      memory_info, initial_tokens.data(), initial_tokens.size(),
+      token_shape.data(), token_shape.size());
+
+  std::array<int64_t, 1> offset_shape{1};
+  Ort::Value offset = Ort::Value::CreateTensor<int64_t>(
+      model_->Allocator(), offset_shape.data(), offset_shape.size());
+  *(offset.GetTensorMutableData<int64_t>()) = 0;
+
+  auto decoder_out = model_->ForwardDecoder(
+      std::move(tokens), std::move(self_kv_cache.first),
+      std::move(self_kv_cache.second), std::move(cross_k), std::move(cross_v),
+      std::move(offset));
+
+  const auto &logits = std::get<0>(decoder_out);
+  const float *p_logits = logits.GetTensorData<float>();
+
+  auto logits_shape = logits.GetTensorTypeAndShapeInfo().GetShape();
+  int32_t vocab_size = logits_shape[2];
+
+  int32_t max_token_id = static_cast<int32_t>(std::distance(
+      p_logits, std::max_element(p_logits, p_logits + vocab_size)));
+
+  int32_t n_text_ctx = model_->TextCtx();
+
+  std::vector<int32_t> predicted_tokens;
+  for (int32_t i = 0; i < n_text_ctx; ++i) {
+    if (max_token_id == model_->EOT()) {
+      break;
+    }
+
+    predicted_tokens.push_back(max_token_id);
+
+    std::array<int64_t, 2> token_shape{1, 1};
+    Ort::Value tokens = Ort::Value::CreateTensor<int64_t>(
+        model_->Allocator(), token_shape.data(), token_shape.size());
+    int64_t *p_tokens = tokens.GetTensorMutableData<int64_t>();
+    p_tokens[0] = max_token_id;
+
+    int64_t *p_offset =
+        std::get<5>(decoder_out).GetTensorMutableData<int64_t>();
+
+    if (i == 0) {
+      *p_offset = initial_tokens.size();
+    } else {
+      *p_offset += 1;
+    }
+
+    decoder_out = model_->ForwardDecoder(std::move(tokens),
+                                         std::move(std::get<1>(decoder_out)),
+                                         std::move(std::get<2>(decoder_out)),
+                                         std::move(std::get<3>(decoder_out)),
+                                         std::move(std::get<4>(decoder_out)),
+                                         std::move(std::get<5>(decoder_out)));
+
+    const auto &logits = std::get<0>(decoder_out);
+    const float *p_logits = logits.GetTensorData<float>();
+
+    max_token_id = static_cast<int64_t>(std::distance(
+        p_logits, std::max_element(p_logits, p_logits + vocab_size)));
+  }
+
+  std::vector<OfflineWhisperDecoderResult> ans(1);
+  ans[0].tokens = std::move(predicted_tokens);
+
+  return ans;
+}
+
+}  // namespace sherpa_onnx
--- a/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h
+++ b/sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h
@@ -0,0 +1,29 @@
+// sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h
+//
+// Copyright (c)  2023  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_GREEDY_SEARCH_DECODER_H_
+#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_GREEDY_SEARCH_DECODER_H_
+
+#include <vector>
+
+#include "sherpa-onnx/csrc/offline-whisper-decoder.h"
+#include "sherpa-onnx/csrc/offline-whisper-model.h"
+
+namespace sherpa_onnx {
+
+class OfflineWhisperGreedySearchDecoder : public OfflineWhisperDecoder {
+ public:
+  explicit OfflineWhisperGreedySearchDecoder(OfflineWhisperModel *model)
+      : model_(model) {}
+
+  std::vector<OfflineWhisperDecoderResult> Decode(Ort::Value cross_k,
+                                                  Ort::Value cross_v) override;
+
+ private:
+  OfflineWhisperModel *model_;  // not owned
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_GREEDY_SEARCH_DECODER_H_
--- a/sherpa-onnx/csrc/offline-whisper-model-config.cc
+++ b/sherpa-onnx/csrc/offline-whisper-model-config.cc
@@ -0,0 +1,46 @@
+// sherpa-onnx/csrc/offline-whisper-model-config.cc
+//
+// Copyright (c)  2023  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/offline-whisper-model-config.h"
+
+#include "sherpa-onnx/csrc/file-utils.h"
+#include "sherpa-onnx/csrc/macros.h"
+
+namespace sherpa_onnx {
+
+void OfflineWhisperModelConfig::Register(ParseOptions *po) {
+  po->Register("whisper-encoder", &encoder,
+               "Path to onnx encoder of whisper, e.g., tiny-encoder.onnx, "
+               "medium.en-encoder.onnx.");
+
+  po->Register("whisper-decoder", &decoder,
+               "Path to onnx decoder of whisper, e.g., tiny-decoder.onnx, "
+               "medium.en-decoder.onnx.");
+}
+
+bool OfflineWhisperModelConfig::Validate() const {
+  if (!FileExists(encoder)) {
+    SHERPA_ONNX_LOGE("whisper encoder file %s does not exist", encoder.c_str());
+    return false;
+  }
+
+  if (!FileExists(decoder)) {
+    SHERPA_ONNX_LOGE("whisper decoder file %s does not exist", decoder.c_str());
+    return false;
+  }
+
+  return true;
+}
+
+std::string OfflineWhisperModelConfig::ToString() const {
+  std::ostringstream os;
+
+  os << "OfflineWhisperModelConfig(";
+  os << "encoder=\"" << encoder << "\", ";
+  os << "decoder=\"" << decoder << "\")";
+
+  return os.str();
+}
+
+}  // namespace sherpa_onnx
--- a/sherpa-onnx/csrc/offline-whisper-model-config.h
+++ b/sherpa-onnx/csrc/offline-whisper-model-config.h
@@ -0,0 +1,30 @@
+// sherpa-onnx/csrc/offline-whisper-model-config.h
+//
+// Copyright (c)  2023  Xiaomi Corporation
+#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
+#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
+
+#include <string>
+
+#include "sherpa-onnx/csrc/parse-options.h"
+
+namespace sherpa_onnx {
+
+struct OfflineWhisperModelConfig {
+  std::string encoder;
+  std::string decoder;
+
+  OfflineWhisperModelConfig() = default;
+  OfflineWhisperModelConfig(const std::string &encoder,
+                            const std::string &decoder)
+      : encoder(encoder), decoder(decoder) {}
+
+  void Register(ParseOptions *po);
+  bool Validate() const;
+
+  std::string ToString() const;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
--- a/sherpa-onnx/csrc/offline-whisper-model.cc
+++ b/sherpa-onnx/csrc/offline-whisper-model.cc
@@ -0,0 +1,213 @@
+// sherpa-onnx/csrc/offline-whisper-model.cc
+//
+// Copyright (c)  2022-2023  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/offline-whisper-model.h"
+
+#include <algorithm>
+#include <string>
+#include <tuple>
+#include <utility>
+
+#include "sherpa-onnx/csrc/macros.h"
+#include "sherpa-onnx/csrc/onnx-utils.h"
+#include "sherpa-onnx/csrc/session.h"
+#include "sherpa-onnx/csrc/text-utils.h"
+
+namespace sherpa_onnx {
+
+class OfflineWhisperModel::Impl {
+ public:
+  explicit Impl(const OfflineModelConfig &config)
+      : config_(config),
+        env_(ORT_LOGGING_LEVEL_ERROR),
+        sess_opts_(GetSessionOptions(config)),
+        allocator_{} {
+    {
+      auto buf = ReadFile(config.whisper.encoder);
+      InitEncoder(buf.data(), buf.size());
+    }
+
+    {
+      auto buf = ReadFile(config.whisper.decoder);
+      InitDecoder(buf.data(), buf.size());
+    }
+  }
+
+  std::pair<Ort::Value, Ort::Value> ForwardEncoder(Ort::Value features) {
+    auto encoder_out = encoder_sess_->Run(
+        {}, encoder_input_names_ptr_.data(), &features, 1,
+        encoder_output_names_ptr_.data(), encoder_output_names_ptr_.size());
+
+    return {std::move(encoder_out[0]), std::move(encoder_out[1])};
+  }
+
+  std::tuple<Ort::Value, Ort::Value, Ort::Value, Ort::Value, Ort::Value,
+             Ort::Value>
+  ForwardDecoder(Ort::Value tokens, Ort::Value n_layer_self_k_cache,
+                 Ort::Value n_layer_self_v_cache, Ort::Value n_layer_cross_k,
+                 Ort::Value n_layer_cross_v, Ort::Value offset) {
+    std::array<Ort::Value, 6> decoder_input = {std::move(tokens),
+                                               std::move(n_layer_self_k_cache),
+                                               std::move(n_layer_self_v_cache),
+                                               std::move(n_layer_cross_k),
+                                               std::move(n_layer_cross_v),
+                                               std::move(offset)};
+
+    auto decoder_out = decoder_sess_->Run(
+        {}, decoder_input_names_ptr_.data(), decoder_input.data(),
+        decoder_input.size(), decoder_output_names_ptr_.data(),
+        decoder_output_names_ptr_.size());
+
+    return {std::move(decoder_out[0]),   std::move(decoder_out[1]),
+            std::move(decoder_out[2]),   std::move(decoder_input[3]),
+            std::move(decoder_input[4]), std::move(decoder_input[5])};
+  }
+
+  std::pair<Ort::Value, Ort::Value> GetInitialSelfKVCache() {
+    std::array<int64_t, 4> shape{n_text_layer_, 1, n_text_ctx_, n_text_state_};
+
+    Ort::Value n_layer_self_k_cache = Ort::Value::CreateTensor<float>(
+        Allocator(), shape.data(), shape.size());
+
+    Ort::Value n_layer_self_v_cache = Ort::Value::CreateTensor<float>(
+        Allocator(), shape.data(), shape.size());
+
+    auto n = shape[0] * shape[1] * shape[2] * shape[3];
+
+    float *p_k = n_layer_self_k_cache.GetTensorMutableData<float>();
+    float *p_v = n_layer_self_v_cache.GetTensorMutableData<float>();
+
+    memset(p_k, 0, sizeof(float) * n);
+    memset(p_v, 0, sizeof(float) * n);
+
+    return {std::move(n_layer_self_k_cache), std::move(n_layer_self_v_cache)};
+  }
+
+  OrtAllocator *Allocator() const { return allocator_; }
+
+  const std::vector<int64_t> &GetInitialTokens() const { return sot_sequence_; }
+
+  int32_t EOT() const { return eot_; }
+
+  int32_t TextCtx() const { return n_text_ctx_; }
+
+ private:
+  void InitEncoder(void *model_data, size_t model_data_length) {
+    encoder_sess_ = std::make_unique<Ort::Session>(
+        env_, model_data, model_data_length, sess_opts_);
+
+    GetInputNames(encoder_sess_.get(), &encoder_input_names_,
+                  &encoder_input_names_ptr_);
+
+    GetOutputNames(encoder_sess_.get(), &encoder_output_names_,
+                   &encoder_output_names_ptr_);
+
+    // get meta data
+    Ort::ModelMetadata meta_data = encoder_sess_->GetModelMetadata();
+    if (config_.debug) {
+      std::ostringstream os;
+      os << "---encoder---\n";
+      PrintModelMetadata(os, meta_data);
+      SHERPA_ONNX_LOGE("%s\n", os.str().c_str());
+    }
+
+    Ort::AllocatorWithDefaultOptions allocator;  // used in the macro below
+    SHERPA_ONNX_READ_META_DATA(n_text_layer_, "n_text_layer");
+    SHERPA_ONNX_READ_META_DATA(n_text_ctx_, "n_text_ctx");
+    SHERPA_ONNX_READ_META_DATA(n_text_state_, "n_text_state");
+    SHERPA_ONNX_READ_META_DATA(sot_, "sot");
+    SHERPA_ONNX_READ_META_DATA(eot_, "eot");
+    SHERPA_ONNX_READ_META_DATA(blank_, "blank_id");
+    SHERPA_ONNX_READ_META_DATA(translate_, "translate");
+    SHERPA_ONNX_READ_META_DATA(no_timestamps_, "no_timestamps");
+    SHERPA_ONNX_READ_META_DATA(no_speech_, "no_speech");
+    SHERPA_ONNX_READ_META_DATA_VEC(sot_sequence_, "sot_sequence");
+  }
+
+  void InitDecoder(void *model_data, size_t model_data_length) {
+    decoder_sess_ = std::make_unique<Ort::Session>(
+        env_, model_data, model_data_length, sess_opts_);
+
+    GetInputNames(decoder_sess_.get(), &decoder_input_names_,
+                  &decoder_input_names_ptr_);
+
+    GetOutputNames(decoder_sess_.get(), &decoder_output_names_,
+                   &decoder_output_names_ptr_);
+  }
+
+ private:
+  OfflineModelConfig config_;
+  Ort::Env env_;
+  Ort::SessionOptions sess_opts_;
+  Ort::AllocatorWithDefaultOptions allocator_;
+
+  std::unique_ptr<Ort::Session> encoder_sess_;
+  std::unique_ptr<Ort::Session> decoder_sess_;
+
+  std::vector<std::string> encoder_input_names_;
+  std::vector<const char *> encoder_input_names_ptr_;
+
+  std::vector<std::string> encoder_output_names_;
+  std::vector<const char *> encoder_output_names_ptr_;
+
+  std::vector<std::string> decoder_input_names_;
+  std::vector<const char *> decoder_input_names_ptr_;
+
+  std::vector<std::string> decoder_output_names_;
+  std::vector<const char *> decoder_output_names_ptr_;
+
+  // model meta data
+  int32_t n_text_layer_;
+  int32_t n_text_ctx_;
+  int32_t n_text_state_;
+  int32_t sot_;
+  int32_t eot_;
+  int32_t blank_;
+  int32_t translate_;
+  int32_t no_timestamps_;
+  int32_t no_speech_;
+  std::vector<int64_t> sot_sequence_;
+};
+
+OfflineWhisperModel::OfflineWhisperModel(const OfflineModelConfig &config)
+    : impl_(std::make_unique<Impl>(config)) {}
+
+OfflineWhisperModel::~OfflineWhisperModel() = default;
+
+std::pair<Ort::Value, Ort::Value> OfflineWhisperModel::ForwardEncoder(
+    Ort::Value features) {
+  return impl_->ForwardEncoder(std::move(features));
+}
+
+std::tuple<Ort::Value, Ort::Value, Ort::Value, Ort::Value, Ort::Value,
+           Ort::Value>
+OfflineWhisperModel::ForwardDecoder(Ort::Value tokens,
+                                    Ort::Value n_layer_self_k_cache,
+                                    Ort::Value n_layer_self_v_cache,
+                                    Ort::Value n_layer_cross_k,
+                                    Ort::Value n_layer_cross_v,
+                                    Ort::Value offset) {
+  return impl_->ForwardDecoder(
+      std::move(tokens), std::move(n_layer_self_k_cache),
+      std::move(n_layer_self_v_cache), std::move(n_layer_cross_k),
+      std::move(n_layer_cross_v), std::move(offset));
+}
+
+std::pair<Ort::Value, Ort::Value> OfflineWhisperModel::GetInitialSelfKVCache() {
+  return impl_->GetInitialSelfKVCache();
+}
+
+OrtAllocator *OfflineWhisperModel::Allocator() const {
+  return impl_->Allocator();
+}
+
+const std::vector<int64_t> &OfflineWhisperModel::GetInitialTokens() const {
+  return impl_->GetInitialTokens();
+}
+
+int32_t OfflineWhisperModel::EOT() const { return impl_->EOT(); }
+
+int32_t OfflineWhisperModel::TextCtx() const { return impl_->TextCtx(); }
+
+}  // namespace sherpa_onnx
--- a/sherpa-onnx/csrc/offline-whisper-model.h
+++ b/sherpa-onnx/csrc/offline-whisper-model.h
@@ -0,0 +1,85 @@
+// sherpa-onnx/csrc/offline-whisper-model.h
+//
+// Copyright (c)  2022-2023  Xiaomi Corporation
+#ifndef SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_H_
+#define SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_H_
+
+#include <memory>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include "onnxruntime_cxx_api.h"  // NOLINT
+#include "sherpa-onnx/csrc/offline-model-config.h"
+
+namespace sherpa_onnx {
+
+class OfflineWhisperModel {
+ public:
+  explicit OfflineWhisperModel(const OfflineModelConfig &config);
+  ~OfflineWhisperModel();
+
+  /** Run the encoder model.
+   *
+   * @param features  A tensor of shape (N, C, T). It is changed in-place.
+   *                  C is 80 and T is 3000.
+   *
+   * @return Return a pair containing:
+   *  - n_layer_cross_k: A 4-D tensor of shape
+   *                     (n_text_layer, N, n_audio_ctx, n_text_state)
+   *  - n_layer_cross_v: A 4-D tensor of shape
+   *                     (n_text_layer, N, n_audio_ctx, n_text_state)
+   */
+  std::pair<Ort::Value, Ort::Value> ForwardEncoder(Ort::Value features);
+
+  /** Run the decoder model.
+   *
+   * @param tokens A int64 tensor of shape (N, num_words)
+   * @param n_layer_self_k_cache  A 4-D tensor of shape
+   *                              (n_text_layer, N, n_text_ctx, n_text_state).
+   * @param n_layer_self_v_cache  A 4-D tensor of shape
+   *                              (n_text_layer, N, n_text_ctx, n_text_state).
+   * @param n_layer_cross_k       A 4-D tensor of shape
+   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
+   * @param n_layer_cross_v       A 4-D tensor of shape
+   *                              (n_text_layer, N, n_audio_ctx, n_text_state).
+   * @param offset A int64 tensor of shape (N,)
+   *
+   * @return Return a tuple containing 6 tensors:
+   *
+   *  - logits A 3-D tensor of shape (N, num_words, vocab_size)
+   *  - out_n_layer_self_k_cache Same shape as n_layer_self_k_cache
+   *  - out_n_layer_self_v_cache Same shape as n_layer_self_v_cache
+   *  - out_n_layer_cross_k Same as n_layer_cross_k
+   *  - out_n_layer_cross_v Same as n_layer_cross_v
+   *  - out_offset Same as offset
+   */
+  std::tuple<Ort::Value, Ort::Value, Ort::Value, Ort::Value, Ort::Value,
+             Ort::Value>
+  ForwardDecoder(Ort::Value tokens, Ort::Value n_layer_self_k_cache,
+                 Ort::Value n_layer_self_v_cache, Ort::Value n_layer_cross_k,
+                 Ort::Value n_layer_cross_v, Ort::Value offset);
+
+  /** Return the initial self kv cache in a pair
+   *  - n_layer_self_k_cache A 4-D tensor of shape
+   *                         (n_text_layer, N, n_audio_ctx, n_text_state).
+   *  - n_layer_self_v_cache A 4-D tensor of shape
+   *                         (n_text_layer, N, n_audio_ctx, n_text_state).
+   */
+  std::pair<Ort::Value, Ort::Value> GetInitialSelfKVCache();
+  const std::vector<int64_t> &GetInitialTokens() const;
+
+  /** Return an allocator for allocating memory
+   */
+  OrtAllocator *Allocator() const;
+  int32_t EOT() const;
+  int32_t TextCtx() const;
+
+ private:
+  class Impl;
+  std::unique_ptr<Impl> impl_;
+};
+
+}  // namespace sherpa_onnx
+
+#endif  // SHERPA_ONNX_CSRC_OFFLINE_WHISPER_MODEL_H_
--- a/sherpa-onnx/csrc/sherpa-onnx-microphone-offline.cc
+++ b/sherpa-onnx/csrc/sherpa-onnx-microphone-offline.cc
@@ -98,11 +98,15 @@ Usage:
  ./bin/sherpa-onnx-microphone-offline \
    --tokens=/path/to/tokens.txt \
    --paraformer=/path/to/model.onnx \
-    --num-threads=2 \
-    --decoding-method=greedy_search
+    --num-threads=1

-Default value for num_threads is 2.
-Valid values for decoding_method: greedy_search.
+(3) Whisper models
+
+  ./bin/sherpa-onnx-microphone-offline \
+    --whisper-encoder=./sherpa-onnx-whisper-base.en/base.en-encoder.int8.onnx \
+    --whisper-decoder=./sherpa-onnx-whisper-base.en/base.en-decoder.int8.onnx \
+    --tokens=./sherpa-onnx-whisper-base.en/base.en-tokens.txt \
+    --num-threads=1

 Please refer to
 https://k2-fsa.github.io/sherpa/onnx/pretrained_models/index.html
--- a/sherpa-onnx/csrc/sherpa-onnx-offline.cc
+++ b/sherpa-onnx/csrc/sherpa-onnx-offline.cc
@@ -23,7 +23,7 @@ Usage:
    --encoder=/path/to/encoder.onnx \
    --decoder=/path/to/decoder.onnx \
    --joiner=/path/to/joiner.onnx \
-    --num-threads=2 \
+    --num-threads=1 \
    --decoding-method=greedy_search \
    /path/to/foo.wav [bar.wav foobar.wav ...]

@@ -33,14 +33,22 @@ Usage:
  ./bin/sherpa-onnx-offline \
    --tokens=/path/to/tokens.txt \
    --paraformer=/path/to/model.onnx \
-    --num-threads=2 \
+    --num-threads=1 \
    --decoding-method=greedy_search \
    /path/to/foo.wav [bar.wav foobar.wav ...]

+(3) Whisper models
+
+  ./bin/sherpa-onnx-offline \
+    --whisper-encoder=./sherpa-onnx-whisper-base.en/base.en-encoder.int8.onnx \
+    --whisper-decoder=./sherpa-onnx-whisper-base.en/base.en-decoder.int8.onnx \
+    --tokens=./sherpa-onnx-whisper-base.en/base.en-tokens.txt \
+    --num-threads=1 \
+    /path/to/foo.wav [bar.wav foobar.wav ...]
+
+
 Note: It supports decoding multiple files in batches

-Default value for num_threads is 2.
-Valid values for decoding_method: greedy_search.
 foo.wav should be of single channel, 16-bit PCM encoded wave file; its
 sampling rate can be arbitrary and does not need to be 16kHz.

@@ -55,6 +63,7 @@ for a list of pre-trained models to download.

  po.Read(argc, argv);
  if (po.NumArgs() < 1) {
+    fprintf(stderr, "Error: Please provide at least 1 wave file.\n\n");
    po.PrintUsage();
    exit(EXIT_FAILURE);
  }
--- a/sherpa-onnx/csrc/symbol-table.cc
+++ b/sherpa-onnx/csrc/symbol-table.cc
@@ -9,6 +9,7 @@
 #include <sstream>
 #include <strstream>

+#include "sherpa-onnx/csrc/base64-decode.h"
 #include "sherpa-onnx/csrc/onnx-utils.h"

 #if __ANDROID_API__ >= 9
@@ -82,4 +83,12 @@ std::ostream &operator<<(std::ostream &os, const SymbolTable &symbol_table) {
  return os << symbol_table.ToString();
 }

+void SymbolTable::ApplyBase64Decode() {
+  sym2id_.clear();
+  for (auto &p : id2sym_) {
+    p.second = Base64Decode(p.second);
+    sym2id_[p.second] = p.first;
+  }
+}
+
 }  // namespace sherpa_onnx
--- a/sherpa-onnx/csrc/symbol-table.h
+++ b/sherpa-onnx/csrc/symbol-table.h
@@ -45,6 +45,9 @@ class SymbolTable {
  /// Return true if there is a given symbol in the symbol table.
  bool contains(const std::string &sym) const;

+  // for tokens.txt from Whisper
+  void ApplyBase64Decode();
+
 private:
  void Init(std::istream &is);

--- a/sherpa-onnx/python/csrc/CMakeLists.txt
+++ b/sherpa-onnx/python/csrc/CMakeLists.txt
@@ -11,6 +11,7 @@ pybind11_add_module(_sherpa_onnx
  offline-recognizer.cc
  offline-stream.cc
  offline-transducer-model-config.cc
+  offline-whisper-model-config.cc
  online-lm-config.cc
  online-recognizer.cc
  online-stream.cc
--- a/sherpa-onnx/python/csrc/offline-model-config.cc
+++ b/sherpa-onnx/python/csrc/offline-model-config.cc
@@ -11,6 +11,7 @@
 #include "sherpa-onnx/python/csrc/offline-nemo-enc-dec-ctc-model-config.h"
 #include "sherpa-onnx/python/csrc/offline-paraformer-model-config.h"
 #include "sherpa-onnx/python/csrc/offline-transducer-model-config.h"
+#include "sherpa-onnx/python/csrc/offline-whisper-model-config.h"

 namespace sherpa_onnx {

@@ -18,22 +19,25 @@ void PybindOfflineModelConfig(py::module *m) {
  PybindOfflineTransducerModelConfig(m);
  PybindOfflineParaformerModelConfig(m);
  PybindOfflineNemoEncDecCtcModelConfig(m);
+  PybindOfflineWhisperModelConfig(m);

  using PyClass = OfflineModelConfig;
  py::class_<PyClass>(*m, "OfflineModelConfig")
-      .def(
-          py::init<const OfflineTransducerModelConfig &,
+      .def(py::init<const OfflineTransducerModelConfig &,
                    const OfflineParaformerModelConfig &,
-                   const OfflineNemoEncDecCtcModelConfig &, const std::string &,
+                    const OfflineNemoEncDecCtcModelConfig &,
+                    const OfflineWhisperModelConfig &, const std::string &,
                    int32_t, bool, const std::string &, const std::string &>(),
           py::arg("transducer") = OfflineTransducerModelConfig(),
           py::arg("paraformer") = OfflineParaformerModelConfig(),
           py::arg("nemo_ctc") = OfflineNemoEncDecCtcModelConfig(),
-          py::arg("tokens"), py::arg("num_threads"), py::arg("debug") = false,
+           py::arg("whisper") = OfflineWhisperModelConfig(), py::arg("tokens"),
+           py::arg("num_threads"), py::arg("debug") = false,
           py::arg("provider") = "cpu", py::arg("model_type") = "")
      .def_readwrite("transducer", &PyClass::transducer)
      .def_readwrite("paraformer", &PyClass::paraformer)
      .def_readwrite("nemo_ctc", &PyClass::nemo_ctc)
+      .def_readwrite("whisper", &PyClass::whisper)
      .def_readwrite("tokens", &PyClass::tokens)
      .def_readwrite("num_threads", &PyClass::num_threads)
      .def_readwrite("debug", &PyClass::debug)
--- a/sherpa-onnx/python/csrc/offline-whisper-model-config.cc
+++ b/sherpa-onnx/python/csrc/offline-whisper-model-config.cc
@@ -0,0 +1,24 @@
+// sherpa-onnx/python/csrc/offline-whisper-model-config.cc
+//
+// Copyright (c)  2023  Xiaomi Corporation
+
+#include "sherpa-onnx/csrc/offline-whisper-model-config.h"
+
+#include <string>
+#include <vector>
+
+#include "sherpa-onnx/python/csrc/offline-whisper-model-config.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineWhisperModelConfig(py::module *m) {
+  using PyClass = OfflineWhisperModelConfig;
+  py::class_<PyClass>(*m, "OfflineWhisperModelConfig")
+      .def(py::init<const std::string &, const std::string &>(),
+           py::arg("encoder"), py::arg("decoder"))
+      .def_readwrite("encoder", &PyClass::encoder)
+      .def_readwrite("decoder", &PyClass::decoder)
+      .def("__str__", &PyClass::ToString);
+}
+
+}  // namespace sherpa_onnx
--- a/sherpa-onnx/python/csrc/offline-whisper-model-config.h
+++ b/sherpa-onnx/python/csrc/offline-whisper-model-config.h
@@ -0,0 +1,16 @@
+// sherpa-onnx/python/csrc/offline-whisper-model-config.h
+//
+// Copyright (c)  2023  Xiaomi Corporation
+
+#ifndef SHERPA_ONNX_PYTHON_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
+#define SHERPA_ONNX_PYTHON_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
+
+#include "sherpa-onnx/python/csrc/sherpa-onnx.h"
+
+namespace sherpa_onnx {
+
+void PybindOfflineWhisperModelConfig(py::module *m);
+
+}
+
+#endif  // SHERPA_ONNX_PYTHON_CSRC_OFFLINE_WHISPER_MODEL_CONFIG_H_
--- a/sherpa-onnx/python/sherpa_onnx/offline_recognizer.py
+++ b/sherpa-onnx/python/sherpa_onnx/offline_recognizer.py
@@ -1,4 +1,5 @@
 # Copyright (c)  2023 by manyeyes
+# Copyright (c)  2023  Xiaomi Corporation
 from pathlib import Path
 from typing import List, Optional

@@ -7,6 +8,7 @@ from _sherpa_onnx import (
    OfflineModelConfig,
    OfflineNemoEncDecCtcModelConfig,
    OfflineParaformerModelConfig,
+    OfflineWhisperModelConfig,
 )
 from _sherpa_onnx import OfflineRecognizer as _Recognizer
 from _sherpa_onnx import (
@@ -69,7 +71,7 @@ class OfflineRecognizer(object):
          feature_dim:
            Dimension of the feature used to train the model.
          decoding_method:
-            Support only greedy_search for now.
+            Valid values: greedy_search, modified_beam_search.
          debug:
            True to show debug messages.
          provider:
@@ -137,7 +139,7 @@ class OfflineRecognizer(object):
          feature_dim:
            Dimension of the feature used to train the model.
          decoding_method:
-            Valid values are greedy_search, modified_beam_search.
+            Valid values are greedy_search.
          debug:
            True to show debug messages.
          provider:
@@ -185,14 +187,14 @@ class OfflineRecognizer(object):
        English, etc.

        Args:
+          model:
+            Path to ``model.onnx``.
          tokens:
            Path to ``tokens.txt``. Each line in ``tokens.txt`` contains two
            columns::

                symbol integer_id

-          model:
-            Path to ``model.onnx``.
          num_threads:
            Number of threads for neural network computation.
          sample_rate:
@@ -200,7 +202,7 @@ class OfflineRecognizer(object):
          feature_dim:
            Dimension of the feature used to train the model.
          decoding_method:
-            Valid values are greedy_search, modified_beam_search.
+            Valid values are greedy_search.
          debug:
            True to show debug messages.
          provider:
@@ -229,6 +231,68 @@ class OfflineRecognizer(object):
        self.recognizer = _Recognizer(recognizer_config)
        return self

+    @classmethod
+    def from_whisper(
+        cls,
+        encoder: str,
+        decoder: str,
+        tokens: str,
+        num_threads: int,
+        decoding_method: str = "greedy_search",
+        debug: bool = False,
+        provider: str = "cpu",
+    ):
+        """
+        Please refer to
+        `<https://k2-fsa.github.io/sherpa/onnx/pretrained_models/index.html>`_
+        to download pre-trained models for different kinds of whisper models,
+        e.g., tiny, tiny.en, base, base.en, etc.
+
+        Args:
+          encoder_model:
+            Path to the encoder model, e.g., tiny-encoder.onnx,
+            tiny-encoder.int8.onnx, tiny-encoder.ort, etc.
+          decoder_model:
+            Path to the encoder model, e.g., tiny-encoder.onnx,
+            tiny-encoder.int8.onnx, tiny-encoder.ort, etc.
+          tokens:
+            Path to ``tokens.txt``. Each line in ``tokens.txt`` contains two
+            columns::
+
+                symbol integer_id
+
+          num_threads:
+            Number of threads for neural network computation.
+          decoding_method:
+            Valid values: greedy_search.
+          debug:
+            True to show debug messages.
+          provider:
+            onnxruntime execution providers. Valid values are: cpu, cuda, coreml.
+        """
+        self = cls.__new__(cls)
+        model_config = OfflineModelConfig(
+            whisper=OfflineWhisperModelConfig(encoder=encoder, decoder=decoder),
+            tokens=tokens,
+            num_threads=num_threads,
+            debug=debug,
+            provider=provider,
+            model_type="whisper",
+        )
+
+        feat_config = OfflineFeatureExtractorConfig(
+            sampling_rate=16000,
+            feature_dim=80,
+        )
+
+        recognizer_config = OfflineRecognizerConfig(
+            feat_config=feat_config,
+            model_config=model_config,
+            decoding_method=decoding_method,
+        )
+        self.recognizer = _Recognizer(recognizer_config)
+        return self
+
    def create_stream(self, contexts_list: Optional[List[List[int]]] = None):
        if contexts_list is None:
            return self.recognizer.create_stream()