Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API. 🧵
Oct 20, 2026 · 4:00 PM UTC
242
376
3,264
270,044
Voxtral Small Transcribe 3 is state-of-the-art in multilingual transcription, diarization and context-biasing.
It detects when someone has finished speaking so agents can respond without delay.
It detects when someone has finished speaking so agents can respond without delay.
11
15
187
16,911
Voxtral Mini TTS 3 achieves state-of-the-art expressiveness in 35+ languages and 50+ dialects.
Implicit instruction-following reads emotion, pace, and emphasis from the conversation.
With explicit instruction-following, tags like [emotion=sad][pace=slow] set them line by line.
Implicit instruction-following reads emotion, pace, and emphasis from the conversation.
With explicit instruction-following, tags like [emotion=sad][pace=slow] set them line by line.
3
4
111
8,517
With Pro Voice, we’re also rolling out professional voice adaptation.
Record up to 30 minutes of audio to replicate your voice with extreme fidelity across every supported language.
Record up to 30 minutes of audio to replicate your voice with extreme fidelity across every supported language.
3
2
81
6,432
Voxtral 3 also introduces a voice agent SDK, built on LiveKit.
Transcribe listens, a reasoning model thinks and calls tools, and Speech replies, streamed as one pipeline for natural, intelligent conversation.
Transcribe listens, a reasoning model thinks and calls tools, and Speech replies, streamed as one pipeline for natural, intelligent conversation.
6
3
74
9,325
@MistralAI 语音 agent API $0.04/分钟，是推理用 Mistral Small 4 的价；换任何其他推理模型是 $0.08/分钟，直接翻倍。另外公告写的是 API 或自有基础设施部署，私有化要找销售谈，没提开放权重下载。
0
0
0
5
@MistralAI Mistral 这次的 Voxtral 3，我觉得最有用的不是转写，是 TTS 能逐句标情绪。拿它给我的 App 推广视频配了段旁白，[像不像真人 / 哪里露馅]。要是真能用，做 UGC 视频又能省一笔配音钱。
0
0
0
206
@MistralAI Pourquoi est-ce que la voix française a l'accent canadien ?
0
0
0
70
@MistralAI Mistral should launch an Android keyboard that has voice input. I would pay to get rid of Google
0
0
11
1,506
@MistralAI the line by line [emotion=sad][pace=slow] tags are the part I'd actually use. are Ukrainian and Czech in the 35+? most TTS I've tried still reads both like a GPS
0
0
2
768
@MistralAI the 6.5% WER on an 1891 phonograph recording is the real number. most speech demos are recorded in a closet with a pop filter
0
0
2
516
@MistralAI Finally, a voice agent that can hear your frustration and continue in 50 dialects.
0
0
6
1,369
@MistralAI ¿cómo anda con audios de WhatsApp en rioplatense, con ruido de colectivo de fondo? ahí se me caían todos los que probé
0
0
0
105
@MistralAI Why does a French AI model have an English accent when speaking French?
0
0
1
129
@MistralAI ¿Pero cuando narices vamos a tener un modelo de voz con esta eficiencia dentro de Vibe? D:
0
0
0
160
@MistralAI A voice that talks in half a second. Mine still sits on hold with customer service.
0
0
0
1
@MistralAI endpointing 这个细节做对了。我之前调语音客服，最大的坑不是识别准不准，是它分不清用户说完没，结果两边一起抢话。转录和语音放一个 API 里，这步挺顺的。
0
0
0
28
@MistralAI @bot summaries this and check the model support Cantonese..?
0
0
0
99
@MistralAI Can you add Finnish to the supported TTS languages?
0
0
0
19