Feature: Use SenseVoice/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.

Hi! Dia's ability to generate realistic dialogue in one pass is impressive.

For preparing training data (audio transcription and labeling), SenseVoice is a natural fit for dialogue TTS:

Key advantages for dialogue TTS data:

• Emotion detection — annotate emotional tone per utterance (happy, sad, angry, neutral)
• Audio event detection — detect laughter, applause, music, coughing
• 5x faster than Whisper — process large datasets quickly
• 50+ languages with strong multilingual support

from funasr import AutoModel

model = AutoModel(model="iic/SenseVoiceSmall")
result = model.generate(input="dialogue.wav")
# Returns text + emotion labels + audio event tags

The built-in emotion recognition is particularly valuable for dialogue TTS, where emotional consistency between text and speech is critical for training quality.

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