[KongchangAI]
ModelVoxtral

Voxtral 3

A family of speech foundation models released by Mistral AI, aimed at real-time speech understanding and generation, with emphasis on low-latency streaming processing, multilingual coverage and open, fine-tunable deployment.

Core Facts

Timeline (last 90 days)

Oct 22

Mistral claims that Voxtral 3 leads Whisper and ElevenLabs in accuracy, but this conclusion comes only from the company's own charts and has no independent benchmark support yet.

Unverified50%
Oct 22

Voxtral 3 pushes the cost of full real-time voice conversation down to 4 cents per minute.

Unverified50%
Oct 22

Voxtral 3 provides a unified voice stack comprising three speech recognition models, one speech generation model, and a real-time voice agent toolkit.

Unverified50%
Oct 22

In early independent developer tests, Voxtral 3's full voice round-trip response was reported to be under one second.

Unverified50%
Oct 22

Voxtral 3's real-time transcription model uses both acoustic signals and semantic information to determine when a caller has finished speaking, thereby reducing awkward pauses and accidental interruptions.

Unverified50%
Oct 22

Voxtral 3 supports more than 50 languages and allows enterprises to write custom terminology directly into the transcription prompt.

Unverified50%
Oct 22

Voxtral 3's batch transcription pricing starts at 0.35 cents per minute.

Unverified50%
Oct 22

Voxtral 3's full real-time agent (paired with Mistral Small 4) is priced at 4 cents per minute; switching to a larger reasoning model doubles the cost to 8 cents per minute.

Unverified50%
Oct 22

Enterprises can run the entire Voxtral 3 stack on European servers or self-host it on private hardware.

Unverified50%
Oct 22

Developers report that Voxtral 3 still requires careful manual audio tuning in noisy environments, and its robustness has not yet reached plug-and-play level.

Unverified50%

10 more timeline events

All Facts (20)

Partially Verified

Mistral AI has released Voxtral 3, the latest iteration in its speech model series.

75%
Unverified

Mistral claims that Voxtral 3 leads Whisper and ElevenLabs in accuracy, but this conclusion comes only from the company's own charts and has no independent benchmark support yet.

50%
Unverified

Voxtral 3 pushes the cost of full real-time voice conversation down to 4 cents per minute.

50%
Unverified

Voxtral 3 provides a unified voice stack comprising three speech recognition models, one speech generation model, and a real-time voice agent toolkit.

50%
Unverified

In early independent developer tests, Voxtral 3's full voice round-trip response was reported to be under one second.

50%
Unverified

Voxtral 3's real-time transcription model uses both acoustic signals and semantic information to determine when a caller has finished speaking, thereby reducing awkward pauses and accidental interruptions.

50%
Unverified

Voxtral 3 supports more than 50 languages and allows enterprises to write custom terminology directly into the transcription prompt.

50%
Unverified

Voxtral 3's batch transcription pricing starts at 0.35 cents per minute.

50%
Unverified

Voxtral 3's full real-time agent (paired with Mistral Small 4) is priced at 4 cents per minute; switching to a larger reasoning model doubles the cost to 8 cents per minute.

50%
Unverified

Enterprises can run the entire Voxtral 3 stack on European servers or self-host it on private hardware.

50%
Unverified

Developers report that Voxtral 3 still requires careful manual audio tuning in noisy environments, and its robustness has not yet reached plug-and-play level.

50%
Unverified

Officially, Voxtral 3 is positioned as state-of-the-art.

50%
Unverified

Officially, only the qualitative statement of state-of-the-art has been given so far; no specific word error rate, MOS score, or first-packet latency data has been disclosed.

50%
Unverified

Voxtral 3 powers Mistral's brand-new voice Agent API.

50%
Unverified

Voxtral 3 is not a single model, but is split into two capability lines: Transcribe (speech-to-text) and Speech (text-to-speech).

50%
Unverified

The original article did not disclose the specific parameters and positioning of the other two transcription models.

50%
Unverified

Voxtral 3 continues Mistral's consistent open-source approach, allowing developers to carry out private deployment and fine-tuning more freely.

50%
Unverified

The specific benchmark data and commercial terms for Voxtral 3 still need to be confirmed by consulting the original announcement.

50%
Unverified

Mistral has launched the public preview of Voxtral 3.

50%
Unverified

The release news of Voxtral 3 received 281 points and 74 comments on Hacker News.

50%

Source Articles