Export KittenTTS mini v0.1 to sherpa-onnx

.github/workflows/export-kitten.yaml

@@ -20,7 +20,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest]
-        version: ["nano_v0_1", "nano_v0_2"]
+        version: ["nano_v0_1", "nano_v0_2", "mini_v0_1"]
         python-version: ["3.10"]
 
     steps:
@@ -59,6 +59,8 @@ jobs:
             d=kitten-nano-en-v0_1-fp16
           elif [[ $version == "nano_v0_2" ]]; then
             d=kitten-nano-en-v0_2-fp16
+          elif [[ $version == "mini_v0_1" ]]; then
+            d=kitten-mini-en-v0_1-fp16
           else
             echo "version $version"
             exit 1
@@ -111,6 +113,7 @@ jobs:
             dirs=(
               kitten-nano-en-v0_1-fp16
               kitten-nano-en-v0_2-fp16
+              kitten-mini-en-v0_1-fp16
             )
 
             export GIT_LFS_SKIP_SMUDGE=1

.github/workflows/generate-tts-samples.yaml

@@ -32,7 +32,7 @@ jobs:
           pip install "numpy<=1.26.4" sherpa-onnx soundfile
 
       - name: kitten
-        if: false
+        if: true
         shell: bash
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
@@ -46,8 +46,9 @@ jobs:
           export GIT_LFS_SKIP_SMUDGE=1
           export GIT_CLONE_PROTECTION_ACTIVE=false
           git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/sherpa-onnx-tts-samples hf
-          mkdir -p ./hf/kitten/v0.1/mp3
-          mkdir -p ./hf/kitten/v0.2/mp3
+          mkdir -p ./hf/kitten/v0.1-nano/mp3
+          mkdir -p ./hf/kitten/v0.2-nano/mp3
+          mkdir -p ./hf/kitten/v0.1-mini/mp3
 
           for v in 1 2; do
             pushd nano_v0_$v
@@ -61,6 +62,18 @@ jobs:
             popd
           done
 
+          for v in 1; do
+            pushd mini_v0_$v
+            curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/kitten-mini-en-v0_$v-fp16.tar.bz2
+            tar xf kitten-mini-en-v0_$v-fp16.tar.bz2
+            rm kitten-mini-en-v0_$v-fp16.tar.bz2
+
+            ln -s ../hf .
+            python3 ./generate_samples.py
+            rm -rf kitten-mini-en-v0_$v-fp16
+            popd
+          done
+
           pushd hf
           git pull
           git add .
@@ -70,7 +83,7 @@ jobs:
           rm -rf hf
 
       - name: matcha en (ljspeech)
-        if: true
+        if: false
         shell: bash
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}

scripts/apk/generate-tts-apk-script.py

@@ -524,6 +524,11 @@ def get_kitten_models() -> List[TtsModel]:
             model_name="model.fp16.onnx",
             lang="en",
         ),
+        TtsModel(
+            model_dir="kitten-mini-en-v0_1-fp16",
+            model_name="model.fp16.onnx",
+            lang="en",
+        ),
     ]
     for m in english_models:
         m.data_dir = f"{m.model_dir}/espeak-ng-data"

scripts/kitten-tts/mini_v0_1/add_meta_data.py

@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+
+import argparse
+
+import numpy as np
+import onnx
+
+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 = np.load("./voices.npz")
+    style_shape = style[list(style.keys())[0]].shape
+
+    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": "kitten-tts",
+        "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/KittenML/kitten-tts-nano-0.2",
+        "see_also": "https://github.com/KittenML/KittenTTS",
+        "maintainer": "k2-fsa",
+        "comment": "This is kitten-tts-nano-0.2 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}")
+
+
+if __name__ == "__main__":
+    main()

scripts/kitten-tts/mini_v0_1/convert_opset.py

@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+"""
+Change the model so that it can be run in onnxruntime 1.17.1
+"""
+
+import onnx
+
+
+def main():
+    model = onnx.load("kitten_tts_mini_v0_1.onnx")
+
+    # Print current opsets
+    for opset in model.opset_import:
+        print(f"Domain: '{opset.domain}', Version: {opset.version}")
+
+    # Modify the opset versions (be careful!)
+    for opset in model.opset_import:
+        if opset.domain == "":  # ai.onnx domain
+            opset.version = 19  # change from 20 to 19
+        elif opset.domain == "ai.onnx.ml":
+            opset.version = 4  # change from 5 to 4
+
+    # Save the modified model
+    onnx.save(model, "model.fp16.onnx")
+
+
+if __name__ == "__main__":
+    main()

scripts/kitten-tts/mini_v0_1/generate_samples.py

@@ -0,0 +1,41 @@
+#!/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(
+        kitten=sherpa_onnx.OfflineTtsKittenModelConfig(
+            model="kitten-mini-en-v0_1-fp16/model.fp16.onnx",
+            voices="kitten-mini-en-v0_1-fp16/voices.bin",
+            tokens="kitten-mini-en-v0_1-fp16/tokens.txt",
+            data_dir="kitten-mini-en-v0_1-fp16/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/kitten/v0.1-mini/mp3/{i}-{s}.mp3",
+        audio.samples,
+        samplerate=audio.sample_rate,
+    )

scripts/kitten-tts/mini_v0_1/generate_tokens.py

@@ -0,0 +1 @@
+../nano_v0_1/generate_tokens.py

scripts/kitten-tts/mini_v0_1/generate_voices_bin.py

@@ -0,0 +1 @@
+../nano_v0_1/generate_voices_bin.py

scripts/kitten-tts/mini_v0_1/run.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+# Copyright      2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+set -ex
+
+if [ ! -f kitten_tts_mini_v0_1.onnx ]; then
+  curl -SL -O https://huggingface.co/KittenML/kitten-tts-mini-0.1/resolve/main/kitten_tts_mini_v0_1.onnx
+fi
+
+if [ ! -f voices.npz ]; then
+  curl -SL -O https://huggingface.co/KittenML/kitten-tts-mini-0.1/resolve/main/voices.npz
+fi
+
+./generate_voices_bin.py
+./generate_tokens.py
+
+./convert_opset.py
+./show.py
+./add_meta_data.py --model ./model.fp16.onnx
+# ./test.py --model ./model.fp16.onnx --tokens ./tokens.txt --voice ./voices.bin
+ls -lh

scripts/kitten-tts/mini_v0_1/show.py

@@ -0,0 +1 @@
+../nano_v0_1/show.py

scripts/kitten-tts/mini_v0_1/test.py

@@ -0,0 +1 @@
+../nano_v0_1/test.py

scripts/kitten-tts/nano_v0_1/generate_samples.py

@@ -35,7 +35,7 @@
     audio = tts.generate(text, sid=i, speed=1.0)
 
     sf.write(
-        f"./hf/kitten/v0.1/mp3/{i}-{s}.mp3",
+        f"./hf/kitten/v0.1-nano/mp3/{i}-{s}.mp3",
         audio.samples,
         samplerate=audio.sample_rate,
     )

scripts/kitten-tts/nano_v0_2/generate_samples.py

@@ -35,7 +35,7 @@
     audio = tts.generate(text, sid=i, speed=1.0)
 
     sf.write(
-        f"./hf/kitten/v0.2/mp3/{i}-{s}.mp3",
+        f"./hf/kitten/v0.2-nano/mp3/{i}-{s}.mp3",
         audio.samples,
         samplerate=audio.sample_rate,
     )