Mistral AI releases Voxtral TTS, its first text-to-speech model

Background

Text-to-speech (TTS) systems turn written text into spoken audio. Early systems assembled speech from recorded fragments; modern ones are neural networks trained on large collections of recorded speech, learning to map characters or phonemes onto acoustic waveforms. Two properties shape how such a model gets used in practice. The first is voice cloning: whether the system can be given a short sample of someone's voice and then read new text in that voice, instead of only offering a fixed set of built-in speakers. The second is latency, the delay between sending text and hearing the first audio. Latency matters most in conversation, where a spoken reply that arrives late breaks the rhythm of dialogue.

What was released

On 23 March 2026, Mistral AI released Voxtral TTS, a text-to-speech model with 4 billion parameters^[1]. The parameter count is the number of learned weights in the network, a rough measure of model size; at 4B the model sits in the range where speech synthesis can run as a single service rather than a large pipeline of separate components.

The model supports 9 languages^[1]. It also supports voice cloning from 3 seconds of audio^[1], meaning a user can supply a very short recording and have new text read in that voice. Three seconds is short enough that a sample can be taken from a phone call or a single sentence of existing material, which is what makes cloning practical for personalised applications such as audiobooks, dubbing, or assistants that speak in a user's own voice.

Licence and pricing

The weights were released under CC BY-NC 4.0^[1]. That licence permits use, sharing and adaptation with attribution, but restricts use to non-commercial purposes, so companies that want to build a product on the model need a separate arrangement. Alongside the open weights, Mistral offers the model through an API priced at $0.016 per 1,000 characters^[1]. Per-character billing ties cost to the length of the text rather than to the length of the resulting audio, which makes budgets easy to estimate for text-heavy uses such as reading articles or documents aloud.

Speed and evaluation

Mistral reported a model latency of 70 ms for a typical 10-second voice sample^[1]. A figure in the tens of milliseconds is far below the roughly one-second scale at which delays start to be noticeable in conversation, so the claim, if it holds in production, puts the model in the range needed for interactive speech applications.

The company also reported that its own human evaluations found Voxtral TTS more natural than ElevenLabs Flash v2.5^[1]. Naturalness is usually judged by asking listeners to rate or compare samples, and it is one of the main quality axes for synthetic speech alongside intelligibility and speaker similarity. The comparison comes from Mistral's internal evaluations rather than an independent benchmark, so it describes the vendor's own measurements.

What it means

The release gives speech output to a model family previously associated with text. It also follows a pattern that has become common for open-weight releases: a downloadable version for research and non-commercial work, and a paid hosted interface for production use, with published per-unit pricing so that developers can compare it against alternatives directly. How the model performs outside Mistral's own tests, on accents, noisy reference samples and languages with less training data, is the kind of question that independent evaluation and wider use will answer.