Mistral's Audio Model Talks Back in 70ms, Becomes Its First Non-Open Model
shashib · x · 2026-08-08
Mistral recently launched Voxtral TTS, its first text-to-speech model, aiming to make voice the primary interface for enterprises to direct AI agents.
Key data points from the analysis:
- 70ms: Claimed model latency for a 10-second voice sample.
- 9.7x: Claimed real-time factor, meaning generated speech arrives significantly faster than the audio's actual length.
- 3 sec: Reference audio required for zero-shot voice cloning.
Notably, Voxtral TTS is also the only model Mistral will not fully open-source, indicating a strategic shift in their licensing approach.
More from Companies & People
- Beff Jezos claims rocket Cybertruck will be the first Tesla/SpaceX lunar vehicle — beffjezos · 2026-10-16
- Debate over why a contested DeepMind researcher is still at the company — lxrjl · 2026-10-16
- Only 5.5% of 400 Indian IT leaders say their firms use an Indian LLM, CLSA survey finds — HarveenChadha · 2026-10-16
- Mistral CEO Arthur Mensch: Europe abandoning frontier AI would mean unprecedented dependence — arthurmensch · 2026-10-16
- Who is Nathan Chen, the Chinese teen AI prodigy courted by Silicon Valley? — scmp_news · 2026-10-16
- From nonprofit to $852B: retelling Sam Altman's OpenAI power struggle — unalteredamoeba · 2026-10-16