Feature: Consider SenseVoice for training data transcription

[!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.

Hi! F5-TTS is great work on flow-matching based speech synthesis.

For training data transcription and annotation, SenseVoice could be useful:

Why SenseVoice for TTS data preparation:

• 5x faster than Whisper large-v3 (234M params, non-autoregressive)
• Emotion detection — classify utterance emotion for expressive TTS training
• Audio event detection — filter out noisy/contaminated audio samples
• 50+ languages — multilingual data annotation
• Word-level timestamps available via OmniSenseVoice

from funasr import AutoModel

model = AutoModel(model="iic/SenseVoiceSmall")
result = model.generate(input="audio.wav")
# Text + emotion + audio events in one pass

GPT-SoVITS (58K stars) already uses FunASR for their training data annotation pipeline with good results.

• FunASR: https://github.com/modelscope/FunASR (16K+ stars)
• SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8K+ stars)