Support GigaAM CTC models for Russian ASR (#1464)

See also https://github.com/salute-developers/GigaAM

scripts/apk/generate-vad-asr-apk-script.py

@@ -333,6 +333,24 @@ def get_models():
 
             ls -lh
 
+            popd
+            """,
+        ),
+        Model(
+            model_name="sherpa-onnx-nemo-ctc-giga-am-russian-2024-10-24",
+            idx=19,
+            lang="ru",
+            short_name="nemo_ctc_giga_am",
+            cmd="""
+            pushd $model_name
+
+            rm -rfv test_wavs
+
+            rm -fv *.sh
+            rm -fv *.py
+
+            ls -lh
+
             popd
             """,
         ),

scripts/nemo/GigaAM/README.md

@@ -0,0 +1,10 @@
+# Introduction
+
+This folder contains scripts for converting models from
+https://github.com/salute-developers/GigaAM
+to sherpa-onnx.
+
+The ASR models are for Russian speech recognition in this folder.
+
+Please see the license of the models at
+https://github.com/salute-developers/GigaAM/blob/main/GigaAM%20License_NC.pdf

scripts/nemo/GigaAM/export-onnx-ctc.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+# Copyright      2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+from typing import Dict
+
+import onnx
+import torch
+import torchaudio
+from nemo.collections.asr.models import EncDecCTCModel
+from nemo.collections.asr.modules.audio_preprocessing import (
+    AudioToMelSpectrogramPreprocessor as NeMoAudioToMelSpectrogramPreprocessor,
+)
+from nemo.collections.asr.parts.preprocessing.features import (
+    FilterbankFeaturesTA as NeMoFilterbankFeaturesTA,
+)
+from onnxruntime.quantization import QuantType, quantize_dynamic
+
+
+class FilterbankFeaturesTA(NeMoFilterbankFeaturesTA):
+    def __init__(self, mel_scale: str = "htk", wkwargs=None, **kwargs):
+        if "window_size" in kwargs:
+            del kwargs["window_size"]
+        if "window_stride" in kwargs:
+            del kwargs["window_stride"]
+
+        super().__init__(**kwargs)
+
+        self._mel_spec_extractor: torchaudio.transforms.MelSpectrogram = (
+            torchaudio.transforms.MelSpectrogram(
+                sample_rate=self._sample_rate,
+                win_length=self.win_length,
+                hop_length=self.hop_length,
+                n_mels=kwargs["nfilt"],
+                window_fn=self.torch_windows[kwargs["window"]],
+                mel_scale=mel_scale,
+                norm=kwargs["mel_norm"],
+                n_fft=kwargs["n_fft"],
+                f_max=kwargs.get("highfreq", None),
+                f_min=kwargs.get("lowfreq", 0),
+                wkwargs=wkwargs,
+            )
+        )
+
+
+class AudioToMelSpectrogramPreprocessor(NeMoAudioToMelSpectrogramPreprocessor):
+    def __init__(self, mel_scale: str = "htk", **kwargs):
+        super().__init__(**kwargs)
+        kwargs["nfilt"] = kwargs["features"]
+        del kwargs["features"]
+        self.featurizer = (
+            FilterbankFeaturesTA(  # Deprecated arguments; kept for config compatibility
+                mel_scale=mel_scale,
+                **kwargs,
+            )
+        )
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """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)
+    while len(model.metadata_props):
+        model.metadata_props.pop()
+
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = str(value)
+
+    onnx.save(model, filename)
+
+
+def main():
+    model = EncDecCTCModel.from_config_file("./ctc_model_config.yaml")
+    ckpt = torch.load("./ctc_model_weights.ckpt", map_location="cpu")
+    model.load_state_dict(ckpt, strict=False)
+    model.eval()
+
+    with open("tokens.txt", "w", encoding="utf-8") as f:
+        for i, t in enumerate(model.cfg.labels):
+            f.write(f"{t} {i}\n")
+        f.write(f"<blk> {i+1}\n")
+
+    filename = "model.onnx"
+    model.export(filename)
+
+    meta_data = {
+        "vocab_size": len(model.cfg.labels) + 1,
+        "normalize_type": "",
+        "subsampling_factor": 4,
+        "model_type": "EncDecCTCModel",
+        "version": "1",
+        "model_author": "https://github.com/salute-developers/GigaAM",
+        "license": "https://github.com/salute-developers/GigaAM/blob/main/GigaAM%20License_NC.pdf",
+        "language": "Russian",
+        "is_giga_am": 1,
+    }
+    add_meta_data(filename, meta_data)
+
+    filename_int8 = "model.int8.onnx"
+    quantize_dynamic(
+        model_input=filename,
+        model_output=filename_int8,
+        weight_type=QuantType.QUInt8,
+    )
+
+
+if __name__ == "__main__":
+    main()

scripts/nemo/GigaAM/run-ctc.sh

@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+# Copyright    2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+set -ex
+
+function install_nemo() {
+  curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
+  python3 get-pip.py
+
+  pip install torch==2.4.0 torchaudio==2.4.0 -f https://download.pytorch.org/whl/torch_stable.html
+
+  pip install -qq wget text-unidecode matplotlib>=3.3.2 onnx onnxruntime pybind11 Cython einops kaldi-native-fbank soundfile librosa
+  pip install -qq ipython
+
+  # sudo apt-get install -q -y sox libsndfile1 ffmpeg python3-pip ipython
+
+  BRANCH='main'
+  python3 -m pip install git+https://github.com/NVIDIA/NeMo.git@$BRANCH#egg=nemo_toolkit[asr]
+
+  pip install numpy==1.26.4
+}
+
+function download_files() {
+  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/ctc_model_weights.ckpt
+  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/ctc_model_config.yaml
+  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/example.wav
+  curl -SL -O https://n-ws-q0bez.s3pd12.sbercloud.ru/b-ws-q0bez-jpv/GigaAM/long_example.wav
+  curl -SL -O https://huggingface.co/csukuangfj/tmp-files/resolve/main/GigaAM%20License_NC.pdf
+}
+
+install_nemo
+download_files
+
+python3 ./export-onnx-ctc.py
+ls -lh
+python3 ./test-onnx-ctc.py

scripts/nemo/GigaAM/test-onnx-ctc.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+# Copyright      2024  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+# https://github.com/salute-developers/GigaAM
+
+import kaldi_native_fbank as knf
+import librosa
+import numpy as np
+import onnxruntime as ort
+import soundfile as sf
+import torch
+
+
+def create_fbank():
+    opts = knf.FbankOptions()
+    opts.frame_opts.dither = 0
+    opts.frame_opts.remove_dc_offset = False
+    opts.frame_opts.preemph_coeff = 0
+    opts.frame_opts.window_type = "hann"
+
+    # Even though GigaAM uses 400 for fft, here we use 512
+    # since kaldi-native-fbank only support fft for power of 2.
+    opts.frame_opts.round_to_power_of_two = True
+
+    opts.mel_opts.low_freq = 0
+    opts.mel_opts.high_freq = 8000
+    opts.mel_opts.num_bins = 64
+
+    fbank = knf.OnlineFbank(opts)
+    return fbank
+
+
+def compute_features(audio, fbank) -> np.ndarray:
+    """
+    Args:
+      audio: (num_samples,), np.float32
+      fbank: the fbank extractor
+    Returns:
+      features: (num_frames, feat_dim), np.float32
+    """
+    assert len(audio.shape) == 1, audio.shape
+    fbank.accept_waveform(16000, audio)
+    ans = []
+    processed = 0
+    while processed < fbank.num_frames_ready:
+        ans.append(np.array(fbank.get_frame(processed)))
+        processed += 1
+    ans = np.stack(ans)
+    return ans
+
+
+def display(sess):
+    print("==========Input==========")
+    for i in sess.get_inputs():
+        print(i)
+    print("==========Output==========")
+    for i in sess.get_outputs():
+        print(i)
+
+
+"""
+==========Input==========
+NodeArg(name='audio_signal', type='tensor(float)', shape=['audio_signal_dynamic_axes_1', 64, 'audio_signal_dynamic_axes_2'])
+NodeArg(name='length', type='tensor(int64)', shape=['length_dynamic_axes_1'])
+==========Output==========
+NodeArg(name='logprobs', type='tensor(float)', shape=['logprobs_dynamic_axes_1', 'logprobs_dynamic_axes_2', 34])
+"""
+
+
+class OnnxModel:
+    def __init__(
+        self,
+        filename: str,
+    ):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.model = ort.InferenceSession(
+            filename,
+            sess_options=session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        display(self.model)
+
+    def __call__(self, x: np.ndarray):
+        # x: (T, C)
+        x = torch.from_numpy(x)
+        x = x.t().unsqueeze(0)
+        # x: [1, C, T]
+        x_lens = torch.tensor([x.shape[-1]], dtype=torch.int64)
+
+        log_probs = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x.numpy(),
+                self.model.get_inputs()[1].name: x_lens.numpy(),
+            },
+        )[0]
+        # [batch_size, T, dim]
+        return log_probs
+
+
+def main():
+    filename = "./model.int8.onnx"
+    tokens = "./tokens.txt"
+    wav = "./example.wav"
+
+    model = OnnxModel(filename)
+
+    id2token = dict()
+    with open(tokens, encoding="utf-8") as f:
+        for line in f:
+            fields = line.split()
+            if len(fields) == 1:
+                id2token[int(fields[0])] = " "
+            else:
+                t, idx = fields
+                id2token[int(idx)] = t
+
+    fbank = create_fbank()
+    audio, sample_rate = sf.read(wav, dtype="float32", always_2d=True)
+    audio = audio[:, 0]  # only use the first channel
+    if sample_rate != 16000:
+        audio = librosa.resample(
+            audio,
+            orig_sr=sample_rate,
+            target_sr=16000,
+        )
+        sample_rate = 16000
+
+    features = compute_features(audio, fbank)
+    print("features.shape", features.shape)
+
+    blank = len(id2token) - 1
+    prev = -1
+    ans = []
+    log_probs = model(features)
+    print("log_probs", log_probs.shape)
+    log_probs = torch.from_numpy(log_probs)[0]
+    ids = torch.argmax(log_probs, dim=1).tolist()
+    for i in ids:
+        if i != blank and i != prev:
+            ans.append(i)
+        prev = i
+
+    tokens = [id2token[i] for i in ans]
+
+    text = "".join(tokens)
+    print(wav)
+    print(text)
+
+
+if __name__ == "__main__":
+    main()