Export kokoro to sherpa-onnx (#1713)

.github/workflows/export-kokoro.yaml

@@ -0,0 +1,122 @@
+name: export-kokoro-to-onnx
+
+on:
+  push:
+    branches:
+      - export-kokoro
+
+  workflow_dispatch:
+
+concurrency:
+  group: export-kokoro-to-onnx-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  export-kokoro-to-onnx:
+    if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
+    name: export kokoro
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.10"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install Python dependencies
+        shell: bash
+        run: |
+          pip install -q "numpy<=1.26.4" onnx==1.16.0 onnxruntime==1.17.1 librosa soundfile piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html
+
+      - name: Run
+        shell: bash
+        run: |
+          curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/espeak-ng-data.tar.bz2
+          tar xf espeak-ng-data.tar.bz2
+          rm espeak-ng-data.tar.bz2
+          cd scripts/kokoro
+          ./run.sh
+
+      - name: Collect results
+        shell: bash
+        run: |
+          src=scripts/kokoro
+
+          d=kokoro-en-v0_19
+          mkdir $d
+          cp -a LICENSE $d/LICENSE
+          cp -a espeak-ng-data $d/
+          cp -v $src/kokoro-v0_19_hf.onnx $d/model.onnx
+          cp -v $src/voices.bin $d/
+          cp -v $src/tokens.txt $d/
+          cp -v $src/README-new.md $d/README.md
+          ls -lh $d/
+          tar cjfv $d.tar.bz2 $d
+          rm -rf $d
+
+          ls -h $.tar.bz2
+
+      - name: Publish to huggingface
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        uses: nick-fields/retry@v3
+        with:
+          max_attempts: 20
+          timeout_seconds: 200
+          shell: bash
+          command: |
+            git config --global user.email "csukuangfj@gmail.com"
+            git config --global user.name "Fangjun Kuang"
+
+            rm -rf huggingface
+            export GIT_LFS_SKIP_SMUDGE=1
+            export GIT_CLONE_PROTECTION_ACTIVE=false
+
+            git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/kokoro-en-v0_19 huggingface
+            cd huggingface
+            rm -rf ./*
+            git fetch
+            git pull
+
+            git lfs track "cmn_dict"
+            git lfs track "ru_dict"
+            git lfs track "*.wav"
+
+            cp -a ../espeak-ng-data ./
+            mkdir -p test_wavs
+
+            cp -v ../scripts/kokoro/kokoro-v0_19_hf.onnx ./model.onnx
+
+            cp -v ../scripts/kokoro/kokoro-v0_19_hf-*.wav ./test_wavs/
+
+            cp -v ../scripts/kokoro/tokens.txt .
+            cp -v ../scripts/kokoro/voices.bin .
+            cp -v ../scripts/kokoro/README-new.md ./README.md
+            cp -v ../LICENSE ./
+
+            git lfs track "*.onnx"
+            git add .
+
+            ls -lh
+
+            git status
+
+            git commit -m "add models"
+            git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/kokoro-en-v0_19 main || true
+
+      - name: Release
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.tar.bz2
+          overwrite: true
+          repo_name: k2-fsa/sherpa-onnx
+          repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
+          tag: tts-models

scripts/kokoro/.gitignore

@@ -0,0 +1,3 @@
+voices.json
+voices.bin
+README-new.md

scripts/kokoro/README.md

@@ -0,0 +1,10 @@
+# Introduction
+
+This folder contains scripts for adding meta data to models
+from https://github.com/thewh1teagle/kokoro-onnx/releases/tag/model-files
+
+See also
+https://huggingface.co/hexgrad/Kokoro-82M/tree/main
+and
+https://huggingface.co/spaces/hexgrad/Kokoro-TTS
+

scripts/kokoro/add-meta-data.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+
+import argparse
+import json
+from pathlib import Path
+
+import numpy as np
+import onnx
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model", type=str, required=True, help="input and output onnx model"
+    )
+
+    parser.add_argument("--voices", type=str, required=True, help="Path to voices.json")
+    return parser.parse_args()
+
+
+def load_voices(filename):
+    with open(filename) as f:
+        voices = json.load(f)
+    for key in voices:
+        voices[key] = np.array(voices[key], dtype=np.float32)
+    return voices
+
+
+def get_vocab():
+    _pad = "$"
+    _punctuation = ';:,.!?¡¿—…"«»“” '
+    _letters = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
+    _letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
+    symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
+    dicts = {}
+    for i in range(len((symbols))):
+        dicts[symbols[i]] = i
+    return dicts
+
+
+def generate_tokens():
+    token2id = get_vocab()
+    with open("tokens.txt", "w", encoding="utf-8") as f:
+        for s, i in token2id.items():
+            f.write(f"{s} {i}\n")
+
+
+def main():
+    args = get_args()
+    print(args.model, args.voices)
+
+    model = onnx.load(args.model)
+    voices = load_voices(args.voices)
+
+    if Path("./tokens.txt").is_file():
+        print("./tokens.txt exist, skip generating it")
+    else:
+        generate_tokens()
+
+    keys = list(voices.keys())
+    print(",".join(keys))
+
+    if Path("./voices.bin").is_file():
+        print("./voices.bin exists, skip generating it")
+    else:
+        with open("voices.bin", "wb") as f:
+            for k in keys:
+                f.write(voices[k].tobytes())
+
+    meta_data = {
+        "model_type": "kokoro",
+        "language": "English",
+        "has_espeak": 1,
+        "sample_rate": 24000,
+        "version": 1,
+        "voice": "en-us",
+        "style_dim": ",".join(map(str, voices[keys[0]].shape)),
+        "n_speakers": len(keys),
+        "speaker_names": ",".join(keys),
+        "model_url": "https://github.com/thewh1teagle/kokoro-onnx/releases/tag/model-files",
+        "see_also": "https://huggingface.co/spaces/hexgrad/Kokoro-TTS",
+        "see_also_2": "https://huggingface.co/hexgrad/Kokoro-82M",
+        "maintainer": "k2-fsa",
+    }
+
+    print(model.metadata_props)
+
+    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)
+    print("--------------------")
+
+    print(model.metadata_props)
+
+    onnx.save(model, args.model)
+
+    print(f"Please see {args.model}, ./voices.bin, and ./tokens.txt")
+
+
+if __name__ == "__main__":
+    main()

scripts/kokoro/run.sh

@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+set -ex
+
+cat > README-new.md <<EOF
+# Introduction
+
+Files in this folder are from
+https://github.com/thewh1teagle/kokoro-onnx/releases/tag/model-files
+
+Please see also
+https://huggingface.co/hexgrad/Kokoro-82M
+and
+https://huggingface.co/hexgrad/Kokoro-82M/discussions/14
+EOF
+
+files=(
+kokoro-v0_19_hf.onnx
+# kokoro-v0_19.onnx
+# kokoro-quant.onnx
+# kokoro-quant-convinteger.onnx
+voices.json
+)
+
+for f in ${files[@]}; do
+  if [ ! -f ./$f ]; then
+    curl -SL -O https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files/$f
+  fi
+done
+
+models=(
+# kokoro-v0_19
+# kokoro-quant
+# kokoro-quant-convinteger
+kokoro-v0_19_hf
+)
+
+for m in ${models[@]}; do
+  ./add-meta-data.py --model $m.onnx --voices ./voices.json
+done
+
+ls -l
+echo "----------"
+ls -lh
+
+for m in ${models[@]}; do
+  ./test.py --model $m.onnx --voices-bin ./voices.bin --tokens ./tokens.txt
+done
+ls -lh

scripts/kokoro/test.py

@@ -0,0 +1,223 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+female (7)
+'af', 'af_bella', 'af_nicole','af_sarah', 'af_sky',
+'bf_emma', 'bf_isabella',
+
+male (4)
+'am_adam',  'am_michael', 'bm_george', 'bm_lewis'
+"""
+
+import argparse
+import time
+from pathlib import Path
+from typing import Dict, List
+
+import numpy as np
+
+try:
+    from piper_phonemize import phonemize_espeak
+except Exception as ex:
+    raise RuntimeError(
+        f"{ex}\nPlease run\n"
+        "pip install piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html"
+    )
+
+import onnxruntime as ort
+import soundfile as sf
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Path to the model",
+    )
+
+    parser.add_argument(
+        "--voices-bin",
+        type=str,
+        required=True,
+        help="Path to the voices.bin",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to tokens.txt",
+    )
+    return parser.parse_args()
+
+
+def show(filename):
+    session_opts = ort.SessionOptions()
+    session_opts.log_severity_level = 3
+    sess = ort.InferenceSession(filename, session_opts)
+    for i in sess.get_inputs():
+        print(i)
+
+    print("-----")
+
+    for i in sess.get_outputs():
+        print(i)
+
+
+#  NodeArg(name='tokens', type='tensor(int64)', shape=[1, 'tokens1'])
+#  NodeArg(name='style', type='tensor(float)', shape=[1, 256])
+#  NodeArg(name='speed', type='tensor(float)', shape=[1])
+#  -----
+#  NodeArg(name='audio', type='tensor(float)', shape=['audio0'])
+
+
+def load_tokens(filename: str) -> Dict[str, int]:
+    ans = dict()
+    with open(filename, encoding="utf-8") as f:
+        for line in f:
+            fields = line.strip().split()
+            if len(fields) == 2:
+                token, idx = fields
+                ans[token] = int(idx)
+            else:
+                assert len(fields) == 1, (len(fields), line)
+                ans[" "] = int(fields[0])
+    return ans
+
+
+def load_voices(speaker_names: List[str], dim: List[int], voices_bin: str):
+    embedding = (
+        np.fromfile(voices_bin, dtype="uint8")
+        .view(np.float32)
+        .reshape(len(speaker_names), *dim)
+    )
+    print("embedding.shape", embedding.shape)
+    ans = dict()
+    for i in range(len(speaker_names)):
+        ans[speaker_names[i]] = embedding[i]
+
+    return ans
+
+
+class OnnxModel:
+    def __init__(self, model_filename: str, voices_bin: str, tokens: str):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+        self.model = ort.InferenceSession(
+            model_filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+        self.token2id = load_tokens(tokens)
+
+        meta = self.model.get_modelmeta().custom_metadata_map
+        print(meta)
+        dim = list(map(int, meta["style_dim"].split(",")))
+        speaker_names = meta["speaker_names"].split(",")
+
+        self.voices = load_voices(
+            speaker_names=speaker_names, dim=dim, voices_bin=voices_bin
+        )
+
+        self.sample_rate = int(meta["sample_rate"])
+
+        print(list(self.voices.keys()))
+        # ['af', 'af_bella', 'af_nicole', 'af_sarah', 'af_sky', 'am_adam',
+        # 'am_michael', 'bf_emma', 'bf_isabella', 'bm_george', 'bm_lewis']
+        # af -> (511, 1, 256)
+        self.max_len = self.voices[next(iter(self.voices))].shape[0] - 1
+
+    def __call__(self, text: str, voice):
+        tokens = phonemize_espeak(text, "en-us")
+        # tokens is List[List[str]]
+        # Each sentence is a List[str]
+        # len(tokens) == number of sentences
+
+        tokens = sum(tokens, [])  # flatten
+        tokens = "".join(tokens)
+
+        tokens = tokens.replace("kəkˈoːɹoʊ", "kˈoʊkəɹoʊ").replace(
+            "kəkˈɔːɹəʊ", "kˈəʊkəɹəʊ"
+        )
+
+        tokens = list(tokens)
+
+        token_ids = [self.token2id[i] for i in tokens]
+        token_ids = token_ids[: self.max_len]
+
+        style = self.voices[voice][len(token_ids)]
+
+        token_ids = [0, *token_ids, 0]
+        token_ids = np.array([token_ids], dtype=np.int64)
+
+        speed = np.array([1.0], dtype=np.float32)
+
+        audio = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: token_ids,
+                self.model.get_inputs()[1].name: style,
+                self.model.get_inputs()[2].name: speed,
+            },
+        )[0]
+        return audio
+
+
+def test(model, voice, text) -> np.ndarray:
+    pass
+
+
+def main():
+    args = get_args()
+    print(vars(args))
+    show(args.model)
+
+    #  tokens = phonemize_espeak("how are you doing?", "en-us")
+    # [['h', 'ˌ', 'a', 'ʊ', ' ', 'ɑ', 'ː', 'ɹ', ' ', 'j', 'u', 'ː', ' ', 'd', 'ˈ', 'u', 'ː', 'ɪ', 'ŋ', '?']]
+    m = OnnxModel(
+        model_filename=args.model, voices_bin=args.voices_bin, tokens=args.tokens
+    )
+
+    text = (
+        "Today as always, men fall into two groups: slaves and free men."
+        + " Whoever does not have two-thirds of his day for himself, "
+        + "is a slave, whatever he may be: a statesman, a businessman, "
+        + "an official, or a scholar."
+    )
+
+    for i, voice in enumerate(m.voices.keys(), 1):
+        print(f"Testing {i}/{len(m.voices)} - {voice}/{args.model}")
+
+        start = time.time()
+        audio = m(text, voice=voice)
+        end = time.time()
+
+        elapsed_seconds = end - start
+        audio_duration = len(audio) / m.sample_rate
+        real_time_factor = elapsed_seconds / audio_duration
+
+        filename = f"{Path(args.model).stem}-{voice}.wav"
+        sf.write(
+            filename,
+            audio,
+            samplerate=m.sample_rate,
+            subtype="PCM_16",
+        )
+        print(f" Saved to {filename}")
+        print(f" Elapsed seconds: {elapsed_seconds:.3f}")
+        print(f" Audio duration in seconds: {audio_duration:.3f}")
+        print(
+            f" RTF: {elapsed_seconds:.3f}/{audio_duration:.3f} = {real_time_factor:.3f}"
+        )
+
+
+if __name__ == "__main__":
+    main()

scripts/melo-tts/export-onnx.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-# This script export ZH_EN TTS model, which supports both Chinese and English.
+# This script exports ZH_EN TTS model, which supports both Chinese and English.
 # This model has only 1 speaker.
 
 from typing import Any, Dict