On March 26, Mistral released Voxtral TTS, its first voice model.
The core selling points: 4B parameters, open weights, human-evaluated naturalness surpassing ElevenLabs Flash v2.5, API pricing at $0.016 per thousand characters, and the ability to run directly on your own server.
How the 4B Parameters Are Built
Voxtral TTS architecture consists of three modules:
| Component | Parameters | Function |
|---|---|---|
| Transformer decoder backbone | 3.4B | Based on Ministral 3B, handles language understanding and text processing |
| Acoustic transformer | 390M | Flow-matching architecture, generates audio features |
| Neural audio codec | 300M | Symmetric codec design, audio compression and reconstruction |
Latency: Given 500 characters of text and 10 seconds of reference audio, the first audio frame comes out within 70 milliseconds. The real-time factor is approximately 9.7x, meaning generating 1 second of speech takes only about 0.1 seconds of compute time.
How It Compares to ElevenLabs
Mistral's announcement states: Human evaluation shows Voxtral TTS voice naturalness exceeds ElevenLabs Flash v2.5 and is on par with ElevenLabs v3 quality.
ElevenLabs is the leading player in the voice AI space, with v3 being their flagship product. The positioning of "on par with flagship, exceeding the sub-flagship" for a 4B open-weight model is a fairly bold claim.
For voice cloning, 3 seconds of reference audio is sufficient. It doesn't simply copy timbre; it captures accent, intonation patterns, pausing habits, and even the rhythm of speech fillers like "uh" and "um." It also supports zero-shot cross-lingual transfer — the same voice configuration can read in different languages while theoretically maintaining consistent accent.
Supported Languages
English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic — 9 languages total.
Chinese is not on the list — this is a significant drawback for Chinese users and means the model can be disregarded for now.
Clarifying the Openness
There's an important detail: Voxtral TTS uses the CC BY NC 4.0 license, not Apache 2.0.
"NC" stands for Non-Commercial — non-commercial use allows free downloading, modification, and distribution, while commercial use requires going through the API. For academic research, personal projects, or content demonstrations, you can download the weights directly from Hugging Face with no usage limits. For integration into commercial products, use the API: $0.016 per thousand characters.
Competitor pricing reference: ElevenLabs Flash v2.5 is approximately $0.066 per thousand characters. Voxtral API is roughly one-quarter of that, offering a clear price advantage.
Who Will Be Affected by This Product
The voice AI landscape has shifted considerably in recent years. ElevenLabs was almost the default choice, then Deepgram, OpenAI, and Microsoft successively added TTS capabilities, while ByteDance's voice products compete domestically in China. Mistral enters with an "open weights + low-cost API" approach — the same strategy it uses for language models.
Voice agents, podcast production, language learning, and customer service systems are the main TTS application scenarios. For enterprises looking to control costs or keep data from leaving their jurisdiction, the self-deployment option is indeed attractive — even for commercial use, you can test with open weights first, confirm the results, then switch to the API.
But to reiterate: Chinese is not supported. This product's primary audience now is European, South Asian, or multilingual overseas teams; there aren't many direct use cases domestically in China.
Sources: Mistral AI just released a text-to-speech model it says beats ElevenLabs (VentureBeat); Mistral releases a new open source model for speech generation (TechCrunch); CocoLoop; Speaking of Voxtral (Mistral AI official blog)