AI Tools & Business / AI news for Malaysia
Mistral Releases Voxtral 3 Speech Models and Voice Agent SDK
Mistral has released four speech models and a toolkit for voice agents, all in Public Preview. The company publishes per-minute prices and Malay among its test languages, while real-world accuracy remains unproven.
In brief
- Mistral’s documentation dates the Voxtral 3 release to 20 October and lists all four models as Public Preview; the announcement page carries a 22 October date.[1][2]
- The family has three transcription models and one text-to-speech model, plus a voice agent SDK built on LiveKit for phone and browser calls.[1][2]
- Mistral publishes per-minute prices and names Malay among its transcription test languages, but its accuracy results are its own and no licence is stated.[1]
What Mistral released
Mistral AI has released Voxtral 3, a family of four speech models and a software kit for building voice agents. The company’s documentation dates the release to 20 October and marks all four models as Public Preview. The announcement page carries a 22 October date.[1][2]
The release covers transcription, live transcription and speech generation. Mistral says its agent kit runs listening, reasoning, tool calls and speaking as one streaming pipeline.[1]
Four models and an agent kit
Voxtral Small Transcribe 3 is the highest-accuracy transcription model, and Voxtral Mini Transcribe 3 handles batch jobs. Both return diarisation, which labels who spoke, along with timestamps. Voxtral Mini Transcribe Realtime 3 streams text during a call, with latency Mistral puts under 500 milliseconds. Voxtral Mini TTS 3 turns text into speech, in batch or in real time.[1][2]
The agent kit joins these models to a general reasoning model, which Mistral names as Mistral Small 4, and to a developer’s own tools. It can take calls over SIP, a standard phone-line protocol, or over WebRTC in a browser or app. Mistral suggests contact centres and service desks as uses but names no customer.[1]
Accuracy claims come from Mistral
Mistral says Voxtral Small outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. In its diarisation test, which measures errors in who said what, lower scores are better. Voxtral Small scored 18.4% and Voxtral Mini 18.6%, against 19.0% for Pyannote Orchestrator and 23.0% for Scribe v2. Gemini 3.5 Transcribe scored 38.4%.[1]
These are Mistral’s charts, and its text-to-speech comparisons used a Gemini model as the judge. They are a reason to run a trial, not a ranking. A noisy call between speakers who switch languages is harder than a benchmark file, and no results on Malaysian recordings are shown.[1]
Prices and availability in preview
Mistral lists batch transcription at US$0.0035 a minute for Mini and US$0.004 for Small. Realtime transcription costs US$0.008 a minute, and text-to-speech costs US$13 per million characters. The voice agent API costs US$0.04 a minute with Mistral Small 4 and US$0.08 a minute with a larger model.[1]
The models run through Mistral’s API, with European, US and global endpoints. On-premises or private-cloud buyers are told to contact sales. The announcement page states no licence, and Public Preview status means terms or limits may still change. Mistral’s notes do not say which ones.[1][2]
Why Malaysia should care
Malay and Indonesian both appear in the 45-language FLEURS benchmark Mistral uses for transcription, and in its 35-language list for realtime transcription. A benchmark list does not show how the models handle everyday Malaysian speech, where callers often switch languages mid-sentence. The realtime list on the page omits Mandarin and Tamil, although Mistral says the model covers 35 or more languages. The announcement does not confirm Malay for text-to-speech. Transcribing calls or building a voice agent are possible uses, not announced deployments.
Contact centres and service desks
Live transcription and an agent kit could make a voice assistant cheaper to prototype at the published per-minute rates.[1]
Practical move: Test recordings of your own calls, including mixed-language speech and background noise.
Developers building voice features
All four models are in Public Preview, so terms, limits or prices could change.[1][2]
Practical move: Prototype against the preview, and keep a fallback provider until terms are final.
Teams handling customer recordings
Mistral lists European, US and global endpoints, and no Malaysian region.[1]
Practical move: Confirm where audio is processed and stored before sending customer calls.
What Malaysians can do now
- Collect a sample of anonymised recordings that reflect your real callers, including mixed-language speech.
- Run the same recordings through your current tool and Voxtral, then compare word errors, speaker labels and response delay.
- Add up the full cost per call, including the reasoning model, retries and your own telephony.
What we still do not know
Mistral has not answered several questions Malaysian buyers may have.
- How accurately the models transcribe Malaysian Malay, English and mixed-language speech, since Mistral publishes benchmark languages rather than Malaysian call data.
- Whether text-to-speech supports Malay, because the announcement does not confirm it.
- Which licence applies, when the preview ends, and whether Malaysian customers can choose where audio is processed.
Sources
- 1.Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages Mistral AI, 22 October 2026
- 2.Changelog Mistral AI Documentation
- 3.File:A home recording setup for the Kathabhidhana project for Wiktionary.jpg Wikimedia Commons, 9 February 2017
- 4.File:Kuala Lumpur. Jalan Petaling. 2019-12-07 15-21-13.jpg Wikimedia Commons, 7 December 2019
- 5.File:Bukit Bintang Junction in 2023.jpg Wikimedia Commons, 26 November 2023
