M5-TTS-50M-ONNX

English | 简体中文

GitHub · Demo · Space · Models

M5-TTS is a compact multilingual and multi-dialect text-to-speech system from ASLP@NPU and LycheeAI Lab. It combines a 30M temporal backbone with a 20M local multi-token prediction decoder, using 16 audio codebooks at 12.5 frames/s for 24 kHz speech generation. The speech codec is separate from the language model.

This repository provides a complete FP32 ONNX inference bundle, including TTS graphs, text/audio embeddings, the text tokenizer and reference-audio encoder/decoder graphs. The same files support CPU and NVIDIA GPU voice cloning; no separate codec download is needed. The source checkpoint is identified in manifest.json and should not be assumed identical to another M5 checkpoint.

Quickstart

Use Python 3.10 on Linux. Install from the code repository:

git clone https://github.com/ASLP-lab/M5-TTS.git
cd M5-TTS
conda create -n m5tts python=3.10 -y
conda activate m5tts
pip install -r requirements.txt

CPU inference, from the repository root:

python infer_onnx.py \
  --model ASLP-lab/M5-TTS-50M-ONNX \
  --device cpu --threads 4 \
  --text "你好，欢迎使用我们的语音合成模型。" \
  --prompt-wav assets/aslp_default_voice.wav \
  --prompt-text "猪携带该疾病，然后通过蚊子传染给人类。" \
  --output outputs/onnx_cpu.wav

NVIDIA GPU inference:

python infer_onnx.py \
  --model ASLP-lab/M5-TTS-50M-ONNX \
  --device cuda --device-id 0 \
  --text "你好，欢迎使用我们的语音合成模型。" \
  --prompt-wav assets/aslp_default_voice.wav \
  --prompt-text "猪携带该疾病，然后通过蚊子传染给人类。" \
  --output outputs/onnx_gpu.wav

CUDA execution requires a compatible NVIDIA driver. The shared requirements install onnxruntime-gpu, which also supports CPU execution; do not additionally install onnxruntime. FlashAttention is not used by this entry point.

Add --stream to write generated audio chunks incrementally. Output is a WAV plus a same-stem JSON statistics file. The current ONNX entry point requires reference audio and its transcript; it does not expose preset speaker IDs or dialect/accent conditioning. CPU and GPU results may differ numerically.

See the ONNX guide and Android example. Only clone voices you own or have permission to use.

License

Apache-2.0. The Qwen-derived audio codec retains its upstream attribution; see the third-party notice.