Add scripts for exporting Piper TTS models to sherpa-onnx (#2299)

scripts/piper/.gitignore

@@ -0,0 +1,5 @@
+*.sh
+*.onnx
+*.json
+MODEL_CARD
+generate_samples-vits-piper*.py

scripts/piper/add_meta_data.py

@@ -0,0 +1,117 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import argparse
+import json
+from typing import Any, Dict
+
+import onnx
+from iso639 import Lang
+
+
+def get_args():
+    # For en_GB-semaine-medium
+    # --name semaine
+    # --kind medium
+    # --lang en_GB
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument(
+        "--name",
+        type=str,
+        required=True,
+    )
+
+    parser.add_argument(
+        "--kind",
+        type=str,
+        required=True,
+    )
+
+    parser.add_argument(
+        "--lang",
+        type=str,
+        required=True,
+    )
+    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)
+
+    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 load_config(filename):
+    with open(filename, "r") as file:
+        config = json.load(file)
+    return config
+
+
+def generate_tokens(config):
+    id_map = config["phoneme_id_map"]
+    with open("tokens.txt", "w", encoding="utf-8") as f:
+        for s, i in id_map.items():
+            f.write(f"{s} {i[0]}\n")
+    print("Generated tokens.txt")
+
+
+# for en_US-lessac-medium.onnx
+# export LANG=en_US
+# export TYPE=lessac
+# export NAME=medium
+def main():
+    args = get_args()
+    print(args)
+    lang = args.lang
+
+    lang_iso = Lang(lang.split("_")[0])
+    print(lang, lang_iso)
+
+    kind = args.kind
+
+    name = args.name
+
+    # en_GB-alan-low.onnx.json
+    config = load_config(f"{lang}-{name}-{kind}.onnx.json")
+
+    print("generate tokens")
+    generate_tokens(config)
+
+    sample_rate = config["audio"]["sample_rate"]
+    if sample_rate == 22500:
+        print("Change sample rate from 22500 to 22050")
+        sample_rate = 22050
+
+    print("add model metadata")
+    meta_data = {
+        "model_type": "vits",
+        "comment": "piper",  # must be piper for models from piper
+        "language": lang_iso.name,
+        "voice": config["espeak"]["voice"],  # e.g., en-us
+        "has_espeak": 1,
+        "n_speakers": config["num_speakers"],
+        "sample_rate": sample_rate,
+    }
+    print(meta_data)
+    add_meta_data(f"{lang}-{name}-{kind}.onnx", meta_data)
+
+
+main()

scripts/piper/dynamic_quantization.py

@@ -0,0 +1,74 @@
+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import argparse
+
+import onnxmltools
+from onnxmltools.utils.float16_converter import convert_float_to_float16
+from onnxruntime.quantization import QuantType, quantize_dynamic
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input",
+        type=str,
+        required=True,
+    )
+    parser.add_argument(
+        "--output-fp16",
+        type=str,
+        required=True,
+    )
+
+    parser.add_argument(
+        "--output-int8",
+        type=str,
+        required=True,
+    )
+    return parser.parse_args()
+
+
+# for op_block_list, see also
+# https://github.com/microsoft/onnxruntime/blob/089c52e4522491312e6839af146a276f2351972e/onnxruntime/python/tools/transformers/float16.py#L115
+#
+# libc++abi: terminating with uncaught exception of type Ort::Exception:
+# Type Error: Type (tensor(float16)) of output arg (/dp/RandomNormalLike_output_0)
+# of node (/dp/RandomNormalLike) does not match expected type (tensor(float)).
+#
+# libc++abi: terminating with uncaught exception of type Ort::Exception:
+# This is an invalid model. Type Error: Type 'tensor(float16)' of input
+# parameter (/enc_p/encoder/attn_layers.0/Constant_84_output_0) of
+# operator (Range) in node (/Range_1) is invalid.
+def export_onnx_fp16(onnx_fp32_path, onnx_fp16_path):
+    onnx_fp32_model = onnxmltools.utils.load_model(onnx_fp32_path)
+    onnx_fp16_model = convert_float_to_float16(
+        onnx_fp32_model,
+        keep_io_types=True,
+        op_block_list=[
+            "RandomNormalLike",
+            "Range",
+        ],
+    )
+    onnxmltools.utils.save_model(onnx_fp16_model, onnx_fp16_path)
+
+
+def main():
+    args = get_args()
+    print(args)
+
+    in_filename = args.input
+    output_fp16 = args.output_fp16
+    output_int8 = args.output_int8
+
+    quantize_dynamic(
+        model_input=in_filename,
+        model_output=output_int8,
+        weight_type=QuantType.QUInt8,
+    )
+
+    export_onnx_fp16(in_filename, output_fp16)
+
+
+if __name__ == "__main__":
+    main()

scripts/piper/generate.sh.in

@@ -0,0 +1,86 @@
+#!/usr/bin/env bash
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# Auto generated! Do NOT edit!
+
+set -ex
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+wget -qq 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
+
+mkdir -p release
+
+{% for model in model_list %}
+
+name={{ model.name }}
+kind={{ model.kind }}
+lang={{ model.lang }}
+model_name={{ model.model_name }}
+text="{{ model.text }}"
+num_speakers={{ model.ns }}
+sample_rate={{ model.sr }}
+
+{{ model.cmd }}
+
+python3 ./add_meta_data.py \
+  --name $name \
+  --kind $kind \
+  --lang $lang
+
+dst=vits-piper-$lang-$name-$kind
+dst_int8=vits-piper-$lang-$name-$kind-int8
+dst_fp16=vits-piper-$lang-$name-$kind-fp16
+mkdir -p $dst
+
+mv -v tokens.txt  $dst/
+mv -v MODEL_CARD $dst/ || true
+mv -v README $dst/ || true
+mv -v *.json  $dst/
+cp -a ./espeak-ng-data $dst/
+
+cp -a $dst $dst_int8
+cp -a $dst $dst_fp16
+
+mv -v *.onnx  $dst/
+
+python3 ./dynamic_quantization.py \
+  --input $dst/$model_name \
+  --output-int8 $dst_int8/$model_name \
+  --output-fp16 $dst_fp16/$model_name >/dev/null 2>&1
+
+echo "---fp32---"
+ls -lh $dst
+
+echo "---int8---"
+ls -lh $dst_int8
+
+echo "---fp16---"
+ls -lh $dst_fp16
+
+tar cjf ${dst}.tar.bz2 $dst
+tar cjf ${dst_int8}.tar.bz2 $dst_int8
+tar cjf ${dst_fp16}.tar.bz2 $dst_fp16
+
+if [ -d hf ]; then
+  mkdir -p hf/piper/mp3/$lang/vits-piper-$lang-$name-$kind
+  for i in $(seq $num_speakers); do
+    i=$((i-1))
+    python3 ./generate_samples-$dst-$i.py
+  done
+  ls -lh hf/piper/mp3/$lang/vits-piper-$lang-$name-$kind
+fi
+
+mv $dst release
+mv $dst_int8 release
+mv $dst_fp16 release
+
+ls -lh release/*
+
+{% endfor %}

scripts/piper/generate_samples.py.in

@@ -0,0 +1,22 @@
+import sherpa_onnx
+import soundfile as sf
+
+config = sherpa_onnx.OfflineTtsConfig(
+    model=sherpa_onnx.OfflineTtsModelConfig(
+        vits=sherpa_onnx.OfflineTtsVitsModelConfig(
+            model="{{ model }}",
+            lexicon="",
+            data_dir="{{ data_dir }}",
+            tokens="{{ tokens }}",
+        ),
+        num_threads=1,
+    ),
+)
+
+if not config.validate():
+    raise ValueError("Please check your config")
+
+tts = sherpa_onnx.OfflineTts(config)
+audio = tts.generate(text="{{text}}", sid={{sid}}, speed=1.0)
+
+sf.write("{{ output_filename }}", audio.samples, samplerate=audio.sample_rate)