Voxtral 3 is a family of speech models and a Voice Agent SDK for real-time conversations. It supports more than 50 languages and combines listening, reasoning, speaking, and tool use into a single process.
This makes it suitable for tasks that require multiple steps during a conversation, such as retrieving information, connecting to business systems, or working through phone and web interfaces. Organizations can choose an API hosted in the European Union or deploy it on their own infrastructure to control how data is managed.
Overview of Voxtral 3
Voxtral 3 is a family of speech models and a voice agent SDK for building real-time conversational systems. It includes both speech-to-text and text-to-speech models within the same family.
Its strengths include speaker diarization, end-of-turn detection, and the ability to adjust emotion, vocal emphasis, and speaking pace according to context. This makes it suitable for contact centers, virtual assistants, and hands-free copilots that need to communicate naturally with users.
When Speech Translation Needs to Lead to Action
Imagine that a user has a referral letter to a dermatologist written in French. The system must read and understand that it is a referral to Dr Martin, then schedule an appointment for Tuesday at 5:45 pm and send the details in English.
This is the real challenge for a Voice Agent. Translation alone is not enough. The system must listen accurately, understand meaning in context, respond naturally, and call the appointment system across multiple steps within the same conversation. If any step is missing, the user still has to complete the task themselves.
Where Voxtral 3 Fits in Mistral’s Model Family
Voxtral 3 is not a single model but Mistral’s new family of speech models. It includes transcription models, speech-generation models, and a Voice Agent SDK that combines listening, reasoning, tool use, and speaking into a single process.
For deployment, organizations can connect through Mistral’s EU-hosted API or install the system on their own infrastructure. The latter provides greater control over where audio is processed, how data is handled, and whether conversations are stored.
From Voxtral Mini Transcribe 2 to the Voxtral 3 Family
The Concatenated Minimum Permutation WER chart shows that Voxtral Small Transcribe 3 scores 18.4% and Voxtral Mini Transcribe 3 scores 18.6%, while Voxtral Mini Transcribe 2 scores 30.9%. Lower is better.
Information about Voxtral Mini Transcribe 2 is limited in the source material, so no conclusions should be drawn about differences in features beyond these test results.
| Factor | Voxtral Mini Transcribe 2 | Voxtral Small Transcribe 3 |
|---|---|---|
| Concatenated Minimum Permutation WER | 30.9% | 18.4% |
| Test results | Test results shown in chart | Test results shown in chart |
How These Features Work in Real-World Situations
Handling Multilingual Conversations and Separating Speakers
In service calls where multiple people interrupt one another, Voxtral Transcribe 3 supports more than 50 languages and can identify who said what, making it easier to follow the conversation accurately. Keyphrases also help the system recognize specialized terminology without requiring fine-tuning.
Knowing When a Speaker Has Finished
End-of-turn detection helps the agent recognize when the other person has finished speaking, allowing it to respond at the right moment. This is useful for contact centers and phone-based assistants that need to hold live conversations.
Adjusting Tone to the Conversation Context
Voxtral Mini TTS 3 adjusts emotion, pacing, and vocal emphasis according to context. It supports more than 35 languages and 50 accents, making spoken responses sound more natural.
Listening, Reasoning, Using Tools, and Continuing the Task
The voice agent SDK combines listening, reasoning, tool use, and speaking in a single pipeline. This allows agents to retrieve information, connect to business systems, and carry out multiple consecutive steps during conversations over the phone or through a website.
Comparing Transcription Results with Other Options
Lower error rates are better. WER is averaged across the 45 languages supported by all seven models, while Japanese, Korean, Thai, and Chinese use CER instead of WER.
| Factor | Voxtral Small Transcribe 3 | Voxtral Mini Transcribe 3 | Whisper | Scribe v2 | Gemini 3.5 Transcribe | Microsoft AI Transcribe 2 |
|---|---|---|---|---|---|---|
| WER (forced language / language detection) | 4.1% / 4.0% | 4.6% / 4.5% | 16.0% / 14.9% | 4.9% / 4.9% | 6.1% / 4.5% | 4.1% / 4.0% |
| Concatenated Minimum Permutation WER | 18.4% | 18.6% | Not available in the chart | 23.0% | 38.4% | 20.3% |
One limitation is that Whisper is not included in the latter chart, and the results are macro-averages based only on the language set specified in the source material.
Strengths and Limitations to Know Before Using It
Voxtral 3 stands out for multilingual listening and response, speaker diarization, context biasing for specialized terminology, and end-of-sentence detection. It is suitable for contact centers and voice assistants that need to perform multiple consecutive steps.
Pros
- +Supports multiple languages, with speaker diarization and end-of-sentence detection
- +Adjusts vocal emotion, connects to multiple tools across several steps, and provides data control
Cons
- −The source material does not specify complete pricing details for every model
- −The test results do not break down every language, and real-world performance depends on the connected systems or tools
Costs and Burdens Not Reflected in the Per-Minute Figure
Voxtral Mini Transcribe 3 lists a price of $0.0035 per minute, but this figure does not cover every usage scenario. Organizations still need to decide whether to use an EU-hosted API or manage the infrastructure for self-deployment themselves.
Another consideration is controlling where audio is processed and how long it is retained. For tasks requiring Pro Voice, users must prepare 30 minutes of voice samples so the system can more precisely adapt the voice characteristics, emotion, and speaking style. The source material does not specify any additional costs for these components.
Who Is It For, and Who Should Wait for More Information?
Voxtral 3 is suitable for contact centers, virtual service providers, and hands-free assistants that need to listen, understand, converse, and perform multiple steps through connected systems. It is well suited to organizations that need control over where audio is processed, how data is managed, and whether conversations are stored.
Teams that need complete real-world pricing for every component or in-depth deployment details should wait for more information before deciding, because the source material does not yet provide these details in full.
The Key Question Is Not Simply Whether the System Understands Speech
Voxtral 3 combines listening, reasoning, tool use, and speaking in a single real-time streaming pipeline. The key issue is therefore not just transcription accuracy, but whether the system can understand the context and complete the task.
When evaluating it, start with a real workflow: receive a request, check information in a system, call the necessary tools, and respond continuously. If all you need is speech-to-text conversion, Voxtral 3 may offer more capability than necessary. But if you need a voice agent that can take action, you need to evaluate the entire process.