Image to help understand the article
Mistral expands its voice AI lineup
Mistral AI unveiled Voxtral 3 on Oct. 20, 2026, introducing a family of models for speech recognition, live transcription, speech generation, voice cloning and voice-based AI agents. The company said its text-to-speech technology supports more than 35 languages.
Batch transcription costs $0.0035 per minute with Voxtral Mini Transcribe 3 and $0.004 per minute with Voxtral Small Transcribe 3. The real-time Voxtral Mini Transcribe Realtime 3 model costs $0.008 per minute. Voxtral Mini TTS 3, the speech-generation model, is priced at $13 per 1 million characters.
Pricing for Mistral’s voice-agent API depends on the underlying model. A configuration using Mistral Small 4 reasoning costs $0.04 per minute, while one using a larger model costs $0.08 per minute.
Available through U.S. and global infrastructure
Voxtral 3 is available through Mistral AI’s European, U.S. and global API endpoints. Organizations seeking to run the technology on their own servers or in a private cloud must contact the company separately. Developers can test transcription, speech generation, voice cloning and voice-agent features in Mistral’s Studio playground.
Mistral said it used Gemini 3.8 Flash as the judging model in an internal evaluation of multilingual and expressive English text-to-speech performance.
Independent transcription results
In a nonstreaming transcription assessment by the independent benchmarking organization Artificial Analysis, Voxtral Small recorded an AA-WER score of 2.1%, while Voxtral Mini scored 2.2%. Lower word error rates indicate more accurate transcription.
Fun-Realtime-ASR-preview led the comparison at 1.7%, followed by MAI-Transcribe-2 at 2.0%. Scribe v2 matched Voxtral Mini at 2.2%, while Gemini 3.5 Transcribe scored 2.6% and GPT Transcribe recorded 3.3%.
The AA-WER benchmark assigns 50% of its score to AA-AgentTalk and 25% each to VoxPopuli-Cleaned-AA and Earnings22-Cleaned-AA.
Comments
Post a Comment