VibeVoice-TTS
by Microsoft · Model · Voice & audio
promising
Small, MIT-licensed speech model from Microsoft, but meant for research only
Paper accepted as an Oral at ICLR 2026; uses 7.5 Hz continuous speech tokenizers with next-token diffusion
AccessOpen weightsNeeds~1.6 GBDownloadabout 1.6 GB (the 1.5B model)LicenseMIT, though Microsoft advises against commercial use
Why we say so
Microsoft reports the paper was accepted as an Oral at ICLR 2026 and the repo has 57.3k stars, which shows interest but not quality. Microsoft itself says it does not recommend commercial use, warns of deepfake misuse and notes biases inherited from its Qwen2.5 base. No independent tests or hands-on reviews are on record.
Evidence
- Nov 18Open-source MIT-licensed Microsoft voice family: TTS-1.5B, Realtime-0.5B and ASR-7B models; repo shows 57.3k stars
- Sep 3VibeVoice-ASR-Streaming released: transcribes while audio arrives, supports hotwords
- Jul 23VibeVoice-ASR-BitNet compresses ASR to 1.58 GB for CPU-only edge inference, real-time on 3+ threads
- Jan 21VibeVoice-ASR-7B released: transcribes 60 minutes in one pass with speaker labels, timestamps, custom hotwords, 50+ languages
- Dec 3, 2025VibeVoice-Realtime-0.5B released: streaming text input, about 300 ms first-audio latency, about 10 minutes of speech
Show all 7
- Sep 5, 2025TTS code removed from the repo after misuse inconsistent with its stated intent; weights page lists the demo as disabled
- Aug 25, 2025Open-sourced long-form multi-speaker TTS: up to 90 minutes in one pass with up to 4 distinct speakers
Cost to run
Alternatives
Made from