Mistral Voxtral TTS: What It Means for Voice AI Developers
March 25, 2026 · 5 min read · Ortavox Team
Mistral just released Voxtral TTS — an open-weights 4B parameter model with 70ms model latency, 3-second voice cloning, and $0.016/1K character pricing. Here is what it changes for developers building voice agents.
On March 23, 2026, Mistral AI released Voxtral TTS — a ~4 billion parameter text-to-speech model that Mistral positions at parity with ElevenLabs v3 in its own listening tests, is cheaper on list price than most hosted alternatives, and is available as open weights on Hugging Face. For developers building voice AI agents, this is a significant development worth paying close attention to.
Update (November 2026): Mistral has since shipped the Voxtral 3 family, including Voxtral Mini TTS 3 (API name voxtral-mini-tts-3-0) at $13 per 1M characters and a realtime TTS variant. The figures below describe the March 2026 Voxtral TTS release as Mistral published them.
What Voxtral TTS is
Voxtral TTS is a transformer-based autoregressive flow-matching model with three components: a 3.4B parameter decoder backbone (built on Ministral 3B), a 390M parameter flow-matching acoustic transformer, and a 300M parameter neural audio codec. Total: approximately 4 billion parameters.
Spec | Value |
|---|---|
Architecture | Transformer + flow-matching acoustic model |
Parameter count | ~4B (3.4B backbone + 390M acoustic + 300M codec) |
Languages | 9 (EN, FR, DE, ES, NL, PT, IT, HI, AR) |
Voice cloning threshold | As little as 3 seconds of reference audio (voice prompt input 5–25 s) |
Model latency | 70ms for a 10-second voice sample and 500-character input (Mistral-reported; not end-to-end time-to-first-audio) |
Real-time factor | ~9.7x (generates audio 9.7x faster than real-time) |
Max output length | Up to 2 minutes natively; longer output handled by the API through interleaving |
Pricing (API) | $0.016 per 1,000 characters ($16 per 1M) |
API / weights | voxtral-mini-tts-2603 on /v1/audio/speech; mistralai/Voxtral-4B-TTS-2603 on Hugging Face |
License | CC BY-NC 4.0 (open weights, non-commercial) |
How it compares to ElevenLabs, OpenAI and Cartesia
Mistral's own listening tests (three annotators, side-by-side preference across nine languages) rated Voxtral more natural than ElevenLabs Flash v2.5 with similar time-to-first-audio, and at parity with ElevenLabs v3 — with emotion-steering support. In voice customization, Mistral reports a 68.4% listener preference for Voxtral over ElevenLabs. These are vendor-run evaluations; no independent benchmark existed at launch. Note that Mistral's 70ms figure is model latency, while the numbers other vendors publish (Cartesia: first audio in under 90ms; ElevenLabs: ~100ms median for Flash v2.5, excluding network) are time-to-first-audio measurements — they are not directly comparable.
Model | Latency (vendor-reported) | List price per 1K chars | Voice cloning | Open weights |
|---|---|---|---|---|
Voxtral TTS (Mistral) | 70ms model latency | $0.016 | Yes (from 3 s of audio) | Yes (CC BY-NC 4.0) |
ElevenLabs Flash v2.5 | ~100ms median (excl. network) | $0.04 (Flash/Turbo API rate) | Yes | No |
ElevenLabs v3 | Not published (v3 Conversational: ~165ms) | $0.08 ($0.04 for v3 Conversational) | Yes | No |
OpenAI tts-1-hd | Not published | $0.03 ($30 per 1M) | No | No |
Cartesia Sonic | First audio in under 90ms | $0.038–$0.065 (plan overage rates) | Yes (from 10 s of audio) | No |
Pricing comparison note: figures are vendor-published list prices (API rates for ElevenLabs and OpenAI, plan overage rates for Cartesia), re-checked in November 2026. ElevenLabs and Cartesia pricing depends on plan, and volume discounts apply on all platforms.
The open-weights angle is significant
Every other major TTS provider in this comparison is fully closed. Voxtral's CC BY-NC 4.0 license means you can download the weights, run inference on your own hardware, and avoid per-character API costs entirely for non-commercial workloads. For deployments with strict data residency requirements or very high call volumes, self-hosting may become economically compelling if a commercial license is available.
The catch: CC BY-NC means no commercial self-hosting without a separate agreement with Mistral. For commercial use, you are still on the API. But the open weights enable fine-tuning, evaluation, and research that closed models cannot.
Voice cloning at 3 seconds
Voxtral's voice adaptation requires as little as 3 seconds of reference audio — Cartesia's Sonic, for comparison, advertises cloning from 10 seconds. More impressively, it supports zero-shot cross-lingual voice adaptation: Mistral's example is English text spoken with a French voice prompt, producing French-accented English. This captures natural personality: pauses, rhythm, intonation, emotional range.
For voice agent applications — where brands often want a consistent agent persona across languages — this is genuinely useful. A single reference voice can be adapted across all 9 supported languages without per-language re-recording.
How to use Voxtral with Ortavox
Ortavox supports Mistral Voxtral as a TTS provider alongside Cartesia, ElevenLabs, OpenAI, Google, Azure and others, and supports bring-your-own-key (BYOK) for LLM, STT and TTS providers. Keys are stored encrypted and never leave the Ortavox API in clear text. To use Voxtral under your own Mistral account, register the key as a credential, then attach it to the agent's TTS configuration:
// POST /v1/credentials — the key is validated against Mistral, encrypted, and stored
{ "provider": "mistral", "name": "Production Mistral", "apiKey": "YOUR_MISTRAL_API_KEY" }
// 201 — only a masked hint is ever returned
{ "id": "cred_123", "provider": "mistral", "name": "Production Mistral", "hint": "1234", "status": "active" }
// Agent TTS settings: pick the Mistral provider and reference the credential
{ "tts": { "provider": "mistralai", "model": "voxtral-mini-tts-latest" }, "ttsCredentialId": "cred_123" }At $0.016/1K chars with v3-class quality claims, Voxtral is one of the cheapest hosted TTS options on list price. Vendor benchmarks are not a substitute for your own: run it against your current ElevenLabs or Cartesia setup on real call audio — especially if you handle multilingual calls.
What this means for the market
Mistral entering TTS puts pricing pressure on ElevenLabs and Cartesia. ElevenLabs' Flash/Turbo API rate of $0.04/1K chars is roughly 2.5x Voxtral's list price, and its v3 rate ($0.08/1K) is 5x. Cartesia's plan overage rates ($38–$65 per 1M characters) land in the same range as Voxtral but without open weights.
For platform-agnostic voice infrastructure — like Ortavox — this is straightforwardly positive. More TTS options at lower cost means cheaper calls for developers. If Voxtral's parity with ElevenLabs v3 holds up in independent testing, the quality trade-off for choosing the cheaper option largely disappears.
Sources
- Mistral AI — Speaking of Voxtral (Voxtral TTS announcement) — https://mistral.ai/news/voxtral-tts/
- Mistral AI — Voxtral 3 announcement — https://mistral.ai/news/voxtral-3
- Mistral AI docs — Audio models overview — https://docs.mistral.ai/en/studio-api/audio/overview
- ElevenLabs docs — Models (latency figures) — https://elevenlabs.io/docs/overview/models
- ElevenLabs — API pricing — https://elevenlabs.io/pricing/api
- OpenAI — API pricing — https://developers.openai.com/api/docs/pricing
- Cartesia — Sonic — https://cartesia.ai/sonic
- Cartesia — Pricing — https://cartesia.ai/pricing
- Ortavox — How it works (supported providers) — https://ortavox.ai/how-it-works