Refactor kokoro export (#2302)

- generate samples for https://k2-fsa.github.io/sherpa/onnx/tts/all/
- provide int8 model for kokoro v0.19 kokoro-int8-en-v0_19.tar.bz2

.github/workflows/export-kokoro.yaml

@@ -3,7 +3,7 @@ name: export-kokoro-to-onnx
 on:
   push:
     branches:
-      - fix-export-kokoro-1.0-2
+      - refactor-kokoro-2
 
   workflow_dispatch:
 
@@ -34,24 +34,94 @@ jobs:
       - name: Install Python dependencies
         shell: bash
         run: |
-          pip install kokoro "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 misaki[en] misaki[zh] torch==2.6.0+cpu -f https://download.pytorch.org/whl/torch
+          pip install kokoro "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 misaki[en] misaki[zh] torch==2.6.0+cpu -f https://download.pytorch.org/whl/torch sherpa-onnx
 
       - name: Run
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         shell: bash
         run: |
           curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/espeak-ng-data.tar.bz2
           tar xf espeak-ng-data.tar.bz2
           rm espeak-ng-data.tar.bz2
+          cp -a ./espeak-ng-data ./scripts/kokoro/v0.19
+          cp -a ./espeak-ng-data ./scripts/kokoro/v1.0
+          cp -a ./espeak-ng-data ./scripts/kokoro/v1.1-zh
+
+          git config --global user.email "csukuangfj@gmail.com"
+          git config --global user.name "Fangjun Kuang"
+
           cd scripts/kokoro
           v=${{ matrix.version }}
           if [[ $v = "0.19" ]]; then
+            cd v0.19
             ./run.sh
+
+            if false; then
+              # generate samples
+              git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-tts-samples hf
+              mkdir -p hf/kokoro/v0.19/mp3
+              ./generate_samples.py
+              pushd hf
+              git pull
+              git add .
+              git commit -m 'add kokoro samples for v0.19'
+              git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-tts-samples main
+              popd
+              rm -rf hf
+            fi
+
           elif [[ $v == "1.0" ]]; then
             cd v1.0
             ./run.sh
+
+            if false; then
+              git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-tts-samples hf
+              mkdir -p hf/kokoro/v1.0/mp3
+
+              curl -SL -O https://github.com/csukuangfj/cppjieba/releases/download/sherpa-onnx-2024-04-19/dict.tar.bz2
+              tar xvf dict.tar.bz2
+              rm dict.tar.bz2
+
+              curl -SL -o date-zh.fst https://huggingface.co/csukuangfj/icefall-tts-aishell3-vits-low-2024-04-06/resolve/main/data/date.fst
+              curl -SL -o number-zh.fst  https://huggingface.co/csukuangfj/icefall-tts-aishell3-vits-low-2024-04-06/resolve/main/data/number.fst
+              curl -SL -o phone-zh.fst https://huggingface.co/csukuangfj/icefall-tts-aishell3-vits-low-2024-04-06/resolve/main/data/phone.fst
+
+              ./generate_samples.py
+              pushd hf
+              git pull
+              git add .
+              git commit -m 'add kokoro samples for v1.0'
+              git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-tts-samples main
+              popd
+              rm -rf hf
+            fi
+
           elif [[ $v == "1.1-zh" ]]; then
             cd v1.1-zh
             ./run.sh
+
+            if false; then
+              git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-tts-samples hf
+              mkdir -p hf/kokoro/v1.1-zh/mp3
+
+              curl -SL -O https://github.com/csukuangfj/cppjieba/releases/download/sherpa-onnx-2024-04-19/dict.tar.bz2
+              tar xvf dict.tar.bz2
+              rm dict.tar.bz2
+
+              curl -SL -o date-zh.fst https://huggingface.co/csukuangfj/icefall-tts-aishell3-vits-low-2024-04-06/resolve/main/data/date.fst
+              curl -SL -o number-zh.fst  https://huggingface.co/csukuangfj/icefall-tts-aishell3-vits-low-2024-04-06/resolve/main/data/number.fst
+              curl -SL -o phone-zh.fst https://huggingface.co/csukuangfj/icefall-tts-aishell3-vits-low-2024-04-06/resolve/main/data/phone.fst
+
+              ./generate_samples.py
+              pushd hf
+              git pull
+              git add .
+              git commit -m 'add kokoro samples for v1.1-zh'
+              git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-tts-samples main
+              popd
+              rm -rf hf
+            fi
           else
             echo "Unknown version $v"
             exit 1
@@ -61,19 +131,39 @@ jobs:
         if: matrix.version == '0.19'
         shell: bash
         run: |
-          src=scripts/kokoro
+          src=scripts/kokoro/v0.19
 
           d=kokoro-en-v0_19
+
           mkdir $d
           cp -a LICENSE $d/LICENSE
           cp -a espeak-ng-data $d/
-          cp -v $src/kokoro-v0_19.onnx $d/model.onnx
+          cp -v $src/model.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
+          cp -v $src/../README.md $d/README.md
+          ls -lh $d/
+          tar cjfv $d.tar.bz2 $d
+
+          ls -lh $d.tar.bz2
+
+      - name: Collect results 0.19 (int8)
+        if: matrix.version == '0.19'
+        shell: bash
+        run: |
+          src=scripts/kokoro/v0.19
+
+          d=kokoro-int8-en-v0_19
+
+          mkdir $d
+          cp -a LICENSE $d/LICENSE
+          cp -a espeak-ng-data $d/
+          cp -v $src/model.int8.onnx $d/model.int8.onnx
+          cp -v $src/voices.bin $d/
+          cp -v $src/tokens.txt $d/
+          cp -v $src/../README.md $d/README.md
           ls -lh $d/
           tar cjfv $d.tar.bz2 $d
-          rm -rf $d
 
           ls -lh $d.tar.bz2
 
@@ -219,33 +309,30 @@ jobs:
             git config --global user.email "csukuangfj@gmail.com"
             git config --global user.name "Fangjun Kuang"
 
-            rm -rf huggingface
+            dirs=(
+              kokoro-en-v0_19
+              # kokoro-int8-en-v0_19
+            )
+
             export GIT_LFS_SKIP_SMUDGE=1
             export GIT_CLONE_PROTECTION_ACTIVE=false
+            for d in ${dirs[@]}; do
+              rm -rf huggingface
 
               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.onnx ./model.onnx
-
-            cp -v ../scripts/kokoro/kokoro-v0_19-*.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 lfs track af_dict
+              git lfs track ar_dict
+              git lfs track cmn_dict
+              git lfs track da_dict en_dict fa_dict hu_dict ia_dict it_dict lb_dict phondata ru_dict ta_dict
+              git lfs track ur_dict yue_dict
+
+
+              cp -a ../$d ./
+
               git add .
 
               ls -lh
@@ -254,6 +341,7 @@ jobs:
 
               git commit -m "add models"
               git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/kokoro-en-v0_19 main || true
+            done
 
       - name: Publish to huggingface 1.0 float32
         if: matrix.version == '1.0'

scripts/kokoro/.gitignore

@@ -1,3 +1,4 @@
+espeak-ng-data
 voices.json
 voices.bin
 README-new.md

scripts/kokoro/README.md

@@ -1,10 +1,6 @@
 # Introduction
 
-This folder contains scripts for adding meta data to models
+Please see also
-from https://github.com/thewh1teagle/kokoro-onnx/releases/tag/model-files
+https://huggingface.co/hexgrad/Kokoro-82M
-
-See also
-https://huggingface.co/hexgrad/Kokoro-82M/tree/main
 and
-https://huggingface.co/spaces/hexgrad/Kokoro-TTS
+https://huggingface.co/hexgrad/Kokoro-82M/discussions/14
-

scripts/kokoro/add_meta_data.py

@@ -1,117 +0,0 @@
-#!/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())
-
-    speaker2id_str = ""
-    id2speaker_str = ""
-    sep = ""
-    for i, s in enumerate(keys):
-        speaker2id_str += f"{sep}{s}->{i}"
-        id2speaker_str += f"{sep}{i}->{s}"
-        sep = ","
-
-    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),
-        "speaker2id": speaker2id_str,
-        "id2speaker": id2speaker_str,
-        "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

@@ -1,50 +0,0 @@
-#!/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/v0.19/.gitignore

@@ -0,0 +1 @@
+kLegacy

scripts/kokoro/v0.19/add_meta_data.py

@@ -0,0 +1,77 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+
+import argparse
+
+import onnx
+import torch
+
+from generate_voices_bin import speaker2id
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model", type=str, required=True, help="input and output onnx model"
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    print(args.model)
+
+    model = onnx.load(args.model)
+
+    style = torch.load(
+        "./kLegacy/v0.19/voices/af.pt", weights_only=True, map_location="cpu"
+    )
+
+    speaker2id_str = ""
+    id2speaker_str = ""
+    sep = ""
+    for s, i in speaker2id.items():
+        speaker2id_str += f"{sep}{s}->{i}"
+        id2speaker_str += f"{sep}{i}->{s}"
+        sep = ","
+
+    meta_data = {
+        "model_type": "kokoro",
+        "language": "English",
+        "has_espeak": 1,
+        "sample_rate": 24000,
+        "version": 1,
+        "voice": "en-us",
+        "style_dim": ",".join(map(str, style.shape)),
+        "n_speakers": len(speaker2id),
+        "speaker2id": speaker2id_str,
+        "id2speaker": id2speaker_str,
+        "speaker_names": ",".join(map(str, speaker2id.keys())),
+        "model_url": "https://huggingface.co/hexgrad/kLegacy/",
+        "see_also": "https://huggingface.co/spaces/hexgrad/Kokoro-TTS",
+        "maintainer": "k2-fsa",
+        "comment": "This is kokoro v0.19 and supports only English",
+    }
+
+    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/v0.19/dynamic_quantization.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+from pathlib import Path
+
+import onnxruntime
+from onnxruntime.quantization import QuantType, quantize_dynamic
+
+
+def show(filename):
+    session_opts = onnxruntime.SessionOptions()
+    session_opts.log_severity_level = 3
+    sess = onnxruntime.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 main():
+    show("./model.onnx")
+
+    if not Path("./model.int8.onnx").is_file():
+        quantize_dynamic(
+            model_input="model.onnx",
+            model_output="model.int8.onnx",
+            #  op_types_to_quantize=["MatMul"],
+            weight_type=QuantType.QUInt8,
+        )
+    else:
+        print("./model.int8.onnx exists - skip")
+
+
+if __name__ == "__main__":
+    main()

scripts/kokoro/v0.19/generate_samples.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+"""
+Generate samples for
+https://k2-fsa.github.io/sherpa/onnx/tts/all/
+"""
+
+import sherpa_onnx
+import soundfile as sf
+
+from generate_voices_bin import speaker2id
+
+config = sherpa_onnx.OfflineTtsConfig(
+    model=sherpa_onnx.OfflineTtsModelConfig(
+        kokoro=sherpa_onnx.OfflineTtsKokoroModelConfig(
+            model="./model.onnx",
+            voices="./voices.bin",
+            tokens="./tokens.txt",
+            data_dir="./espeak-ng-data",
+        ),
+        num_threads=2,
+    ),
+    max_num_sentences=1,
+)
+
+if not config.validate():
+    raise ValueError("Please check your config")
+
+tts = sherpa_onnx.OfflineTts(config)
+text = "Friends fell out often because life was changing so fast. The easiest thing in the world was to lose touch with someone."
+
+for s, i in speaker2id.items():
+    print(s, i, len(speaker2id))
+    audio = tts.generate(text, sid=i, speed=1.0)
+
+    sf.write(
+        f"./hf/kokoro/v0.19/mp3/{i}-{s}.mp3",
+        audio.samples,
+        samplerate=audio.sample_rate,
+    )

scripts/kokoro/v0.19/generate_tokens.py

@@ -0,0 +1,26 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+
+def get_vocab():
+    # https://huggingface.co/hexgrad/kLegacy/blob/main/v0.19/kokoro.py#L75
+    _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 main():
+    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")
+
+
+if __name__ == "__main__":
+    main()

scripts/kokoro/v0.19/generate_voices_bin.py

@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+import torch
+from pathlib import Path
+
+
+id2speaker = {
+    0: "af",
+    1: "af_bella",
+    2: "af_nicole",
+    3: "af_sarah",
+    4: "af_sky",
+    5: "am_adam",
+    6: "am_michael",
+    7: "bf_emma",
+    8: "bf_isabella",
+    9: "bm_george",
+    10: "bm_lewis",
+}
+
+speaker2id = {speaker: idx for idx, speaker in id2speaker.items()}
+
+
+def main():
+    if Path("./voices.bin").is_file():
+        print("./voices.bin exists - skip")
+        return
+
+    with open("voices.bin", "wb") as f:
+        for _, speaker in id2speaker.items():
+            m = torch.load(
+                f"kLegacy/v0.19/voices/{speaker}.pt",
+                weights_only=True,
+                map_location="cpu",
+            ).numpy()
+            # m.shape (511, 1, 256)
+
+            f.write(m.tobytes())
+
+
+if __name__ == "__main__":
+    main()

scripts/kokoro/v0.19/run.sh

@@ -0,0 +1,36 @@
+#!/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
+git clone https://huggingface.co/hexgrad/kLegacy
+EOF
+
+if [ ! -d kLegacy ]; then
+  git clone https://huggingface.co/hexgrad/kLegacy
+  pushd kLegacy/v0.19
+  git lfs pull
+  popd
+fi
+
+if [ ! -f ./voices.bin ]; then
+  ./generate_voices_bin.py
+fi
+
+if [ ! -f ./tokens.txt ]; then
+  ./generate_tokens.py
+fi
+
+if [ ! -f ./model.onnx ]; then
+  mv kLegacy/v0.19/kokoro-v0_19.onnx ./model.onnx
+fi
+
+./add_meta_data.py --model ./model.onnx
+
+if [ ! -f model.int8.onnx ]; then
+  ./dynamic_quantization.py
+fi

scripts/kokoro/v0.19/test.py

@@ -67,11 +67,13 @@ def show(filename):
         print(i)
 
 
-#  NodeArg(name='tokens', type='tensor(int64)', shape=[1, 'tokens1'])
+"""
-#  NodeArg(name='style', type='tensor(float)', shape=[1, 256])
+NodeArg(name='tokens', type='tensor(int64)', shape=[1, 'tokens1'])
-#  NodeArg(name='speed', type='tensor(float)', shape=[1])
+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'])
+-----
+NodeArg(name='audio', type='tensor(float)', shape=['audio0'])
+"""
 
 
 def load_tokens(filename: str) -> Dict[str, int]:
@@ -171,10 +173,6 @@ class OnnxModel:
         return audio
 
 
-def test(model, voice, text) -> np.ndarray:
-    pass
-
-
 def main():
     args = get_args()
     print(vars(args))

scripts/kokoro/v1.0/generate_samples.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+"""
+Generate samples for
+https://k2-fsa.github.io/sherpa/onnx/tts/all/
+"""
+
+import sherpa_onnx
+import soundfile as sf
+
+from generate_voices_bin import speaker2id
+
+config = sherpa_onnx.OfflineTtsConfig(
+    model=sherpa_onnx.OfflineTtsModelConfig(
+        kokoro=sherpa_onnx.OfflineTtsKokoroModelConfig(
+            model="./kokoro.onnx",
+            voices="./voices.bin",
+            tokens="./tokens.txt",
+            data_dir="./espeak-ng-data",
+            dict_dir="./dict",
+            lexicon="./lexicon-zh.txt,./lexicon-us-en.txt",
+        ),
+        num_threads=2,
+        debug=True,
+    ),
+    rule_fsts="./phone-zh.fst,./date-zh.fst,./number-zh.fst",
+    max_num_sentences=1,
+)
+
+if not config.validate():
+    raise ValueError("Please check your config")
+
+tts = sherpa_onnx.OfflineTts(config)
+text = "This model supports both Chinese and English. 小米的核心价值观是什么？答案是真诚热爱！有困难，请拨打110 或者18601200909。I am learning 机器学习. 我在研究 machine learning。What do you think 中英文说的如何呢? 今天是 2025年6月18号."
+
+print("text", text)
+
+for s, i in speaker2id.items():
+    print(s, i, len(speaker2id))
+    audio = tts.generate(text, sid=i, speed=1.0)
+
+    sf.write(
+        f"./hf/kokoro/v1.0/mp3/{i}-{s}.mp3",
+        audio.samples,
+        samplerate=audio.sample_rate,
+    )

scripts/kokoro/v1.1-zh/generate_samples.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+"""
+Generate samples for
+https://k2-fsa.github.io/sherpa/onnx/tts/all/
+"""
+
+import sherpa_onnx
+import soundfile as sf
+
+from generate_voices_bin import speaker2id
+
+config = sherpa_onnx.OfflineTtsConfig(
+    model=sherpa_onnx.OfflineTtsModelConfig(
+        kokoro=sherpa_onnx.OfflineTtsKokoroModelConfig(
+            model="./kokoro.onnx",
+            voices="./voices.bin",
+            tokens="./tokens.txt",
+            data_dir="./espeak-ng-data",
+            dict_dir="./dict",
+            lexicon="./lexicon-zh.txt,./lexicon-us-en.txt",
+        ),
+        num_threads=2,
+        debug=True,
+    ),
+    rule_fsts="./phone-zh.fst,./date-zh.fst,./number-zh.fst",
+    max_num_sentences=1,
+)
+
+if not config.validate():
+    raise ValueError("Please check your config")
+
+tts = sherpa_onnx.OfflineTts(config)
+text = "This model supports both Chinese and English. 小米的核心价值观是什么？答案是真诚热爱！有困难，请拨打110 或者18601200909。I am learning 机器学习. 我在研究 machine learning。What do you think 中英文说的如何呢? 今天是 2025年6月18号."
+
+print("text", text)
+
+for s, i in speaker2id.items():
+    print(s, i, len(speaker2id))
+    audio = tts.generate(text, sid=i, speed=1.0)
+
+    sf.write(
+        f"./hf/kokoro/v1.1-zh/mp3/{i}-{s}.mp3",
+        audio.samples,
+        samplerate=audio.sample_rate,
+    )

scripts/kokoro/v1.1-zh/run.sh

@@ -11,6 +11,8 @@ fi
 if [ ! -f config.json ]; then
   # see https://huggingface.co/hexgrad/Kokoro-82M/blob/main/config.json
   curl -SL -O https://huggingface.co/hexgrad/Kokoro-82M-v1.1-zh/resolve/main/config.json
+  mkdir -p Kokoro-82M
+  cp ./config.json ./Kokoro-82M
 fi
 
 voices=(

sherpa-onnx/python/csrc/offline-tts.cc

@@ -34,7 +34,7 @@ static void PybindOfflineTtsConfig(py::module *m) {
       .def(py::init<const OfflineTtsModelConfig &, const std::string &,
                     const std::string &, int32_t, float>(),
            py::arg("model"), py::arg("rule_fsts") = "",
-           py::arg("rule_fars") = "", py::arg("max_num_sentences") = 2,
+           py::arg("rule_fars") = "", py::arg("max_num_sentences") = 1,
            py::arg("silence_scale") = 0.2)
       .def_readwrite("model", &PyClass::model)
       .def_readwrite("rule_fsts", &PyClass::rule_fsts)