VibeVoice — распознавание и синтез речи

★ 52.8K

VibeVoice is Microsoft's open-source voice framework covering both sides of audio work: speech recognition and speech synthesis. The VibeVoice-ASR-7B recognition model processes up to 60 minutes of audio in a single pass and returns a transcript with speaker separation — useful for recorded calls, interviews and podcasts where who said what matters. The synthesis side does the opposite: it speaks text aloud and holds a consistent voice character across long passages, which is what audiobooks and multi-speaker dialogue need. Reach for it when you need both directions in one stack rather than wiring together separate projects, and when processing must stay on your own hardware with no recordings leaving the building. It runs locally; a GPU is required and memory needs depend on the model you choose.