Feature Request: Use FunASR for training data annotation

[!NOTE] License and capability clarification (2026-07-14): FunASR is a toolkit, not a single checkpoint. The FunASR and SenseVoice repository source code is MIT; model weights follow each model card. SenseVoiceSmall supports Chinese, Cantonese, English, Japanese, and Korean, and its weights use the linked FunASR Model Open Source License Agreement. Fun-ASR-Nano-2512 is Apache-2.0. Language coverage, punctuation, and performance depend on the selected model and runtime configuration.

Feature Request

MeloTTS is an excellent multi-lingual TTS project! FunASR could help with training data preparation — automatic speech-to-text annotation for audio datasets.

Precedent: GPT-SoVITS (58K stars) already uses FunASR for training data labeling.

Why FunASR for TTS data prep?

• SenseVoice: 50+ languages including Chinese, English, Japanese, Korean — matching MeloTTS's language coverage
• Paraformer: Accurate character-level timestamps for audio-text alignment
• Built-in VAD: Automatic audio segmentation
• Built-in punctuation: Automatic punctuation restoration
• Fast: 170x realtime on GPU — efficient for large dataset annotation

Example:

from funasr import AutoModel
model = AutoModel(model="paraformer-zh", vad_model="fsmn-vad", punc_model="ct-punc")
result = model.generate(input="training_audio.wav")
# Returns: timestamped, punctuated text segments

• GitHub: https://github.com/modelscope/FunASR (16K+ stars)