{"id":9936,"date":"2026-10-26T15:18:20","date_gmt":"2026-10-26T15:18:20","guid":{"rendered":"https:\/\/intnews.it\/?p=9936"},"modified":"2026-10-26T15:18:20","modified_gmt":"2026-10-26T15:18:20","slug":"ai-voice-agents-voxtral-3","status":"publish","type":"post","link":"https:\/\/intnews.it\/en\/ai-voice-agents-voxtral-3\/","title":{"rendered":"AI Voice Agents: How Mistral’s Voxtral 3 Works"},"content":{"rendered":"\n
<\/div>\n\n\n\n

An AI voice agent is software that answers the phone, understands what the caller wants, acts on business systems and replies out loud. On 20 October 2026 Mistral AI released Voxtral 3, a family of speech models plus a software development kit (SDK) for building such agents.<\/p>\n\n\n\n

Voxtral 3 is not a finished product for a shop or a clinic. It is a set of building blocks that developers and software vendors assemble into phone assistants, billed by the minute. Mistral’s pitch is sovereignty: the stack runs on its EU-hosted API or on a company’s own infrastructure, so the business controls where call audio is processed and whether it is kept.<\/p>\n\n\n\n

Last updated: October 2026<\/strong><\/p>\n\n\n\n

<\/div>\n\n\n\n

How does a Voxtral 3 voice agent work?<\/h2>\n\n\n\n

The SDK is built on LiveKit, an open-source platform for real-time audio and video. It chains three models into one streaming pipeline:<\/p>\n\n\n\n

    \n
  1. Listen.<\/strong> Voxtral Mini Transcribe Realtime 3 transcribes the call as the person speaks and detects when they have finished.<\/li>\n\n\n\n
  2. Reason and act.<\/strong> A language model, such as Mistral Small, interprets the request and calls connected tools: calendars, business software, document stores, web search.<\/li>\n\n\n\n
  3. Reply.<\/strong> Voxtral Mini TTS 3 speaks, adjusting tone, pace and emphasis to the conversation.<\/li>\n<\/ol>\n\n\n\n

    No stage waits for the previous one to finish. Mistral says this keeps each conversational turn under one second and lets the agent stop talking when the caller interrupts. These are vendor figures, not independent measurements.<\/p>\n\n\n\n

    During a call the agent can verify a caller’s identity, check a bill or book an appointment. It connects to phone systems over SIP, the standard protocol for internet telephony, and to websites and apps over WebRTC. Because each stage is a separate model, a business can swap the voice, the language or the language model without rebuilding the agent.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    How much does a Voxtral 3 voice agent cost?<\/h2>\n\n\n\n

    Mistral sells the individual models and a complete agent service.<\/p>\n\n\n\n

    Component<\/td>Model<\/td>What it does<\/td>List price<\/td><\/tr><\/thead>
    Full agent<\/td>Voice agent API with Mistral Small 4<\/td>Listening, reasoning and speech in one service<\/td>$0.04 per minute<\/td><\/tr>
    Full agent<\/td>Voice agent API with a larger model<\/td>Same service, more capable language model<\/td>$0.08 per minute<\/td><\/tr>
    Real-time listening<\/td>Voxtral Mini Transcribe Realtime 3<\/td>Live call transcription, claimed latency under 500 ms<\/td>$0.008 per minute<\/td><\/tr>
    File transcription<\/td>Voxtral Mini Transcribe 3<\/td>Recorded audio<\/td>$0.0035 per minute<\/td><\/tr>
    High-accuracy transcription<\/td>Voxtral Small Transcribe 3<\/td>Most accurate model in the family<\/td>$0.004 per minute<\/td><\/tr>
    Synthetic voice<\/td>Voxtral Mini TTS 3<\/td>Spoken replies<\/td>$13 per million characters<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    Source: Mistral AI price list, 20 October 2026.<\/em><\/p>\n\n\n\n

    A three-minute call on the base agent costs $0.12; 200 such calls a month come to $24. The price covers the models, not the phone line: connecting a number over SIP takes a telephony carrier with its own charges. The API is served from European, US and global endpoints; running the full stack on-premises goes through Mistral’s sales team.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    How accurate is Voxtral 3 on the phone?<\/h2>\n\n\n\n

    There is an independent benchmark for transcription. On the Artificial Analysis leaderboard checked on 26 October 2026, Voxtral Small Transcribe 3 ranks third of 59 models, with 2.1% of words transcribed incorrectly. Alibaba’s Fun-Realtime-ASR-preview (1.7%) and Microsoft’s MAI-Transcribe-2 (2.0%) rank ahead of it; the Mini version is fifth at 2.2%.<\/p>\n\n\n\n

    That leaderboard measures transcription of recorded files. Phone calls run on the real-time model, and Mistral’s own tests across 35 languages show the trade-off: a 6.4% word error rate with half a second of delay and 6.0% with about one second, against 4.0% for file transcription on the same test. The less time the model waits, the less context it has and the more it gets wrong.<\/p>\n\n\n\n

    Language coverage also depends on the model. File transcription reaches 50 languages, including all 24 official EU languages. The real-time model covers more than 35 (the model card on the same Mistral page says more than 40). The synthetic voice speaks more than 35 languages and over 50 regional variants. Mistral also publishes quality-per-dollar comparisons for its voice, but they rest on internal evaluations.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    What do EU rules require from AI phone agents?<\/h2>\n\n\n\n

    A business deploying a voice agent in the EU already works under these rules.<\/p>\n\n\n\n

    Rule<\/td>What it requires<\/td>Applies from<\/td>Who it binds<\/td><\/tr><\/thead>
    EU AI Act (Regulation 2024\/1689), Article 50<\/td>Tell callers they are talking to an AI, clearly and at the latest at the first interaction, unless it is obvious<\/td>2 August 2026<\/td>The provider of the system, including a company that builds its own agent and runs it under its own name<\/td><\/tr>
    GDPR<\/td>A notice and a legal basis for recording or transcribing calls<\/td>25 May 2018<\/td>Whoever handles the calls<\/td><\/tr>
    Italian Criminal Code, Article 612-quater (Law 132\/2025)<\/td>1 to 5 years in prison for sharing, without consent, AI-faked images, video or voices that can deceive and cause unjust harm<\/td>10 October 2025<\/td>Anyone, in Italy<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    On a phone line, the Article 50 disclosure in practice belongs at the start of the call. The EU’s Digital Omnibus postponed other AI Act obligations, but not this one.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    What is Italy doing about voice cloning?<\/h2>\n\n\n\n

    Italy is the first EU country with a comprehensive national AI law. Law 132\/2025, in force since 10 October 2025, added Article 612-quater to the Criminal Code. The offence is not creating a synthetic voice: it is handing over, publishing or spreading one without consent when it can deceive and causes unjust harm. Prosecution normally requires a complaint from the victim, with exceptions, for instance when the target is a public authority attacked for their role.<\/p>\n\n\n\n

    On 5 October 2026 it emerged that Prime Minister Giorgia Meloni had filed with the EUIPO, the EU’s intellectual property office, a four-second recording in which she says “Io sono Giorgia Meloni” (“I am Giorgia Meloni”) twice. It is an application for a sound mark, still under examination according to reports at the time; the singer Giusy Ferreri and the voice actor Luca Ward took the same route earlier. A sound mark is legal protection, not a technical one: it helps against unauthorised commercial use, but it does not stop anyone from cloning a voice.<\/p>\n\n\n\n

    That matters because Voxtral 3 includes Pro Voice, which Mistral says reproduces a speaker’s character, emotional range and style, across languages, from 30 minutes of audio. Mistral’s announcement does not explain how the consent of the person whose voice is reproduced is verified.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    What does this mean for small businesses?<\/h2>\n\n\n\n

    A hair salon or a dental practice will not use Voxtral 3 directly. It will use a vertical service built for its trade from components like these, and falling per-minute prices matter most to the companies building those services. Voice agents are the phone-based branch of the AI agents<\/a> now taking on tasks at work.<\/p>\n\n\n\n

    SalonVoice, an Italian company founded by Christian La Porta, is one example. It is not a Mistral product, and its website does not say which voice models it uses. According to the company, a salon forwards its calls to a SalonVoice number when nobody can answer. The assistant reads services and staff from the salon’s management software, writes the booking into the calendar and puts unusual requests on a call-back list. It also tells callers it is an automated assistant.<\/p>\n\n\n\n

    The pricing logic differs from Mistral’s: the salon pays for outcomes, not minutes. SalonVoice charges \u20ac20 a month including 10 bookings, then \u20ac1.50 per booking. The company reports 72 bookings in three months at one salon in a shopping centre in the province of Monza, with the assistant active from 7 pm to 9:30 am between 24 June and 25 September 2026. That is a single case measured by the vendor: it shows the mechanism works, not what it earns on average. For salons, it adds to the uses of AI covered in our beauty tech<\/a> guide.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    Frequently asked questions about AI voice agents<\/h2>\n\n\n\n
    What is an AI voice agent?<\/strong>

    Software that answers the phone, understands the request, acts on business systems and replies out loud. It can book an appointment, check a bill or verify a caller’s identity.<\/p> <\/div>

    What is Voxtral 3?<\/strong>

    Mistral AI’s family of speech models released on 20 October 2026: transcription, a synthetic voice and an SDK for building voice agents.<\/p> <\/div>

    How much does a Voxtral 3 voice agent cost?<\/strong>

    According to Mistral’s price list, $0.04 per minute with Mistral Small 4 and $0.08 with a larger model. Phone-line charges are extra.<\/p> <\/div>

    Which languages does Voxtral 3 support?<\/strong>

    File transcription covers 50 languages, including the 24 official EU languages. The real-time model used on calls covers more than 35, and the synthetic voice more than 35 plus regional variants.<\/p> <\/div>

    Do businesses in the EU have to disclose that an AI is answering?<\/strong>

    Yes. Since 2 August 2026, Article 50 of the AI Act requires callers to be told at the latest at the first interaction, unless it is obvious.<\/p> <\/div>

    Is voice cloning a crime in Italy?<\/strong>

    Cloning itself is not. Article 612-quater of the Italian Criminal Code punishes handing over, publishing or spreading an AI-faked voice without consent when it can deceive and causes unjust harm.<\/p> <\/div><\/div>\n\n\n\n

    <\/div>\n\n\n\n

    INT News event<\/h2>\n\n\n\n

    INT News’s Interconnection evening in Seregno, north of Milan, on 26 October 2026 (6:30 to 8:30 pm) is about voice agents: Christian La Porta will show SalonVoice answering calls and booking appointments for a real business. Details and free registration on the event page<\/a>.<\/p>\n\n\n\n

    <\/div>\n\n\n\n

    Sources<\/h2>\n\n\n\n