Mistral Just Open-Sourced a TTS Model That Beats ElevenLabs

90 milliseconds. That's how fast Voxtral starts talking

On March 26, Mistral AI released Voxtral, its first text-to-speech model. It's a 4-billion-parameter model, available as open weights on Hugging Face under a CC BY NC 4.0 license. For non-commercial use, it's free.

Here's why this matters: in blind human evaluations, Voxtral was rated more natural-sounding than ElevenLabs Flash v2.5. ElevenLabs is the de facto standard in voice AI, a company that just closed a $500M Series D at an $11B valuation.

Mistral gave away a model that beats that for free.

The backstory — Why this disrupts the voice AI market

AI-powered TTS has been around for years, but the "sounds like a real person" threshold was crossed around 2023-2024. ElevenLabs led that transition with high-quality voice cloning, multilingual support, and emotional expression, all behind a paid API.

The problem is cost. When enterprises embed ElevenLabs into customer support bots, per-minute charges add up fast at scale.

Feature	ElevenLabs v3	ElevenLabs Flash v2.5	Voxtral TTS
Model size	Proprietary	Proprietary	4B parameters
TTFA (500 chars)	Proprietary	Approx. 90ms	90ms
Naturalness (human eval)	Top tier	High	Higher than Flash v2.5
Pricing	$0.024/1K chars	$0.008/1K chars	API $0.016/1K chars, self-host free
Min voice clone sample	10-30s	10-30s	Under 5 seconds
Languages	29	29	9
License	Proprietary	Proprietary	CC BY NC 4.0 (open weight)

Voxtral doesn't fully replace ElevenLabs. It supports only 9 languages, and commercial use requires a separate license. But the key insight is this:

Self-host it and the marginal cost is zero. You just need a GPU.

What makes Voxtral technically different

Voice cloning from under 5 seconds

Voxtral's killer feature is ultra-short voice cloning. Less than five seconds of audio is enough to capture accents, intonation, speech patterns, and even irregularities like stuttering. And those voice characteristics persist across all 9 supported languages.

In practice, this means a single short greeting from a CEO can power a consistent voice identity across a 9-language customer support system.

6x real-time factor

An RTF of 6x means a 10-second clip renders in approximately 1.6 seconds. That's fast enough for real-time conversational AI agents, where the full cycle of user question, AI processing, and voice response needs to feel like natural conversation.

Emotion steering

Voxtral supports emotion control. The same text delivered with "joyful," "serious," or "comforting" labels produces noticeably different speech styles. Evaluators rated this capability at parity with ElevenLabs v3.

The bigger picture — Open source is coming for voice AI

Voxtral signals that the pattern we saw in text LLMs is repeating in voice.

In 2024, Meta's Llama series proved open-source LLMs could compete with proprietary ones. Mistral, Qwen, and DeepSeek followed with competitive open models, giving enterprises alternatives to API-only dependencies.

The same pattern is now playing out in voice. Voxtral could be voice AI's "Llama moment" — an open model matching commercial quality.

Mistral raising $830M in debt the same week to buy 13,800 Nvidia chips and expand its Paris data center fits the picture. Open-source the model, monetize through API service and enterprise licenses. It's the hybrid playbook that's proving out across AI.

What this means for you

If you're a developer, try it now. Search mistralai/Voxtral-4B-TTS-2603 on Hugging Face. At 4B parameters, it runs on consumer GPUs like the RTX 4090.

If you're running a voice AI service, it's time to re-examine your cost structure. For services in the 9 supported languages (English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic), self-hosted Voxtral could dramatically cut your ElevenLabs API costs.

One caveat: commercial use requires a separate license from Mistral. CC BY NC 4.0 covers non-commercial use only.

---

References

• Mistral releases a new open source model for speech generation — TechCrunch
• Speaking of Voxtral — Mistral AI
• Mistral AI just released a text-to-speech model it says beats ElevenLabs — VentureBeat
• mistralai/Voxtral-4B-TTS-2603 — Hugging Face