VibeVoice

Why we included this project

VibeVoice pairs a long-form ASR model with a multi-speaker TTS line, so one repo covers both directions of the speech pipeline. The ASR side transcribes up to 60 minutes of audio in a single pass, returns speaker labels and timestamps, and accepts custom hotwords for domain terms, which makes it useful for meeting transcription or call analytics rather than just demos. It runs natively in over 50 languages, and the finetuning code plus a quantized CPU inference engine (no GPU required) cover the path from training to edge deployment. The TTS side handles long-form, multi-speaker synthesis with streaming input. Models are on Hugging Face, with Colab demos and vLLM integration, so teams can go from download to a working pipeline quickly.