Voxtral 3: create an AI voice agent with Mistral

Mistral Voxtral 3 makes it possible to create an AI voice agent capable of listening, understanding, querying business tools, and responding orally. Launched on October 20, 2026, the lineup covers file transcription in 50 languages, real-time transcription in 45 languages and dialects, and voice generation in 35 languages. For a small or midsize business, usage costs can remain low, but telephone integration, security, and business testing determine the project’s success.

What does Mistral Voxtral 3 contribute to an AI voice agent?

Mistral Voxtral 3 is a family of five audio models that transforms speech into text or text into voice. Its Voice Agents SDK, built on LiveKit, combines these functions with a language model, business tools, and continuous audio streaming. Callers can interrupt the agent while it is responding.

The lineup published by Mistral in 2026 includes three speech recognition models and two speech synthesis models. The references are voxtral-mini-transcribe-3-0, voxtral-small-transcribe-3-0, voxtral-mini-transcribe-realtime-3-0, voxtral-mini-tts-3-0 and voxtral-mini-tts-realtime-3-0.

An AI voice agent is software that conducts a orale conversation and can trigger controlled actions in other applications. The Voice Agents SDK captures audio, transcribes it, sends the text to a configurable language model, autorizes tool calls, generates the voice response, and then sends the audio back to the user.

This separation matters to decision-makers. You can change the model responsible for reasoning without rebuilding the entire audio layer, or limit the tools available according to the scenario. To place the conversational component within the provider’s ecosystem, our Mistral guide for SMBs and GDPR explains the related governance issues.

How do you create a phone system with Mistral Voxtral 3?

A phone system based on Mistral Voxtral 3 requires more than an API connection. You must connect the telephone system to the Voice Agents SDK, convert the audio stream, configure the language model, control autorized actions, and provide for human intervention. Mistral’s documentation requires the SDK to receive 16-bit mono PCM audio at 48 kHz.

Telephones often transmit a signal at 8 kHz. A custom audio interface or telephone gateway must therefore receive the call and convert this signal to the expected format. This is a subtle pitfall: a successful browser demonstration does not prove that quality will be identical on a telephone line.

For an appointment-scheduling service, the integration flow can follow these steps:

1. The gateway receives the call, informs the caller that they are interacting with an automated system, and transmits a converted audio stream.
2. voxtral-mini-transcribe-realtime-3-0 produces the text as the conversation progresses.
3. The language model determines the intent and, if necessary, requests the execution of a business tool.
4. The client’s application checks available time slots and explicitly confirms the details before recording the appointment.
5. The speech synthesis model delivers the response, while a signed webhook signals the beginning, end, or failure of the call.

Custom tools run in the company’s application, not automatically at Mistral. The hosted tools offered by Mistral may include web search, document libraries, and workspace connectors. From an agency perspective, the best practice is to limit each tool to a narrowly defined action: reading a calendar does not autorize modifying an entire customer record.

Human intervention should be planned for sensitive requests, repeated misunderstandings, and irreversible actions. Frankly, allowing an agent to confirm a cancellation or payment on its own from the first version creates a disproportionate risk compared with the expected benefit.

How much does a Mistral Voxtral 3 voice agent cost in 2026?

The official Mistral Voxtral 3 API prices published in 2026 start at $0.0035 per minute for Mini batch transcription. Real-time transcription costs $0.008 per minute, and voice generation costs $13 per million characters. These amounts exclude the language model, telephone service, hosting, and storage.

A scenario involving 1,000 calls lasting 4 minutes represents 4,000 minutes of real-time transcription, or $32 at Mistral’s 2026 rates. With approximately 1,500 characters generated per call, synthesis adds $19.50. The total audio cost therefore reaches approximately $51.50, before language model reasoning, the telephone carrier, the servers, monitoring, and archiving.

The API price is generally not the largest cost item in a pilot project. Dialogue planning, connecting to business software, testing on noisy calls, and securing actions require more work. At this level of consumption, it is better to invest in failure scenarios than to save a few dollars by choosing a different transcription model.

What latency and accuracy should you expect in real-world conditions?

In 2026, Mistral Voxtral 3’s real-time transcription offers a configurable delay of 80 to 1,200 milliseconds, as well as a 2,400-millisecond setting. Reducing this delay makes the interaction more responsive but decreases accuracy. An agent’s total response time also includes reasoning, tools, synthesis, and network latency.

The launch evaluations reported by Mistral report, on FLEURS-35, a WER of 5.7 % with a delay of 2,400 ms and 7.4 % at 240 ms. WER, or word error rate, measures the proportion of words that are substituted, deleted, or added. These results come from the provider and do not measure a complete call.

An independent evaluation reported by Artificial Analysis on October 21, 2026, assigned a WER of 2.1 % to Voxtral Small Transcribe 3 and 2.2 % to Voxtral Mini Transcribe 3. These figures concern transcription, not voice quality, tool reliability, or end-to-end latency.

One day after the lorunch, no published independent test yet covered all 35 synthesis languages, the 45 real-time transcription varieties, tool calls, or long voice responses. In the projects we lead, we therefore test representative recordings: real accents, product names, office noise, interruptions, and telephone-quality audio.

Also pay attention to the language figures. In 2026, Mistral announced 50 languages for files, 45 languages and dialects in real time, then 35 languages and more than 50 dialects for generated speech. A language understood by transcription is not necessarily available with an appropriate generated voice.

How can you secure data and choose hosting?

Mistral Voxtral 3 can use API endpoints hosted in the European Union, the United States, or globally. Mistral also announced full deployment on a private cloud or on premises through its sales team. The choice should depend on the data processed, the contracts, and the expected level of control.

A telephone conversation may contain an identity, a phone number, an address, or sensitive data. The RGPD, applicable since 2018, notably requires a specific purpose, a legal basis, understandable information, and a proportionate retention period. Clearly inform the caller about the automation and, if the audio is recorded, about the purpore reserved for that recording.

In 2026, Mistral stated that organizations can control where audio is processed and conversations are retained. However, the exact terms should not be inferred from a sales page: review the service agreement, the data processing accorrd, the subprocessors, and the settings applied to the account.

Tool calls require separate protection. Use minimum permissions, validate parameters on the server side, log actions, and require confirmation before any high-impact operation. Signed webhooks help authenticate start, completion, or failure notifications, but they do not replace an access-control policy.

Also plan for the unavailability of the provider or telephone gateway. A transfer to a human, a fallback message, and a continuity procedure prevent an outage from blocking the phone system. Preparing a cloud incident recovery plan and monitoring the data entrusted to a service provider provide useful guidelines for this framework.

Define a voice agent before its development prevents most surprises related to telephony, access rights, and human takeover. An outside perspective can especially help transform a convincing demonstration into an operational, measurable, and reversible service.

FAQ about Mistral Voxtral 3

Can Mistral Voxtral 3 speak French?

Mistral Voxtral 3 supports voice generation in 35 languages in 2026, while more than 50 dialects have been announced. Quality must be tested using the vocabulary, proper names, and telephone conditions specific to the company.

Does Voxtral 3 work directly with a phone line?

Voxtral 3 requires a custom audio interface or gateway to receive a phone call. In particular, an 8 kHz telephone signal must be converted to 48 kHz, 16-bit mono PCM audio for the Voice Agents SDK.

Can Voxtral 3 be connected to a CRM or calendar?

The Voice Agents SDK authorizes custom tools executed within the company’s application. A connector can access a CRM or calendar, provided that permissions are restricted, data is validated, and actions are logged.

Can Voxtral 3 be hosted in Europe?

In 2026, Mistral will offer an API endpoint hosted in the European Union, as well as options in the United States or globally. Deployment on a private cloud or on-premises will also be available through Mistral’s sales team.