Mistral Releases Voxtral TTS: 4B Parameter Text-to-Speech Model at $0.016 per 1k Characters
Mistral AI has released Voxtral TTS, a 4B parameter text-to-speech model supporting 9 languages including English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic. The model achieves 70ms latency for typical inputs and can clone voices from as little as 3 seconds of audio, priced at $0.016 per 1,000 characters.
Mistral Releases Voxtral TTS: 4B Parameter Text-to-Speech Model at $0.016 per 1k Characters
Mistral AI has released Voxtral TTS, a 4B parameter text-to-speech model supporting 9 languages with voice cloning capabilities from as little as 3 seconds of audio.
Technical Specifications
The model consists of three components:
- 3.4B parameter transformer decoder backbone (built on Ministral 3B)
- 390M parameter flow-matching acoustic transformer
- 300M parameter neural audio codec
Voxtral TTS achieves 70ms model latency for typical inputs (10-second voice sample, 500 characters) with a real-time factor of approximately 9.7x. The model natively generates up to 2 minutes of audio, with the API handling longer generations through smart interleaving.
Language Support and Capabilities
Voxtral TTS supports 9 languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic. The model can adapt to custom voices using reference samples as short as 3 seconds, capturing voice characteristics including accent, inflections, intonations, and disfluencies.
The model demonstrates zero-shot cross-lingual voice adaptation despite not being explicitly trained for it. For example, it can generate English speech using a French voice prompt, producing natural-sounding French-accented English.
Architecture Details
Voxtral TTS uses a transformer-based, autoregressive, flow-matching architecture. The in-house codec processes audio causally using semantic VQ (8,192 vocabulary) and acoustic FSQ (36 dimensions, 21 levels) latent representations, producing them at 12.5Hz frame rate. The flow-matching transformer runs 16 function evaluations per audio frame to produce acoustic latents.
Performance Benchmarks
According to Mistral AI, human evaluations by native speakers show Voxtral TTS achieves superior naturalness compared to ElevenLabs Flash v2.5 while maintaining similar time-to-first-audio. The company claims performance parity with ElevenLabs v3 in quality, with support for emotion-steering.
In zero-shot custom voice evaluations across 9 languages, Mistral conducted side-by-side preference tests with 3 annotators per language pair, measuring naturalness, accent adherence, and acoustic similarity. The company claims Voxtral TTS outperformed ElevenLabs v2.5 Flash in this multilingual custom voice setting.
Pricing and Availability
Voxtral TTS is available via API at $0.016 per 1,000 characters. The model can be tested in Mistral Studio and Le Chat. An open-weight version with reference voices is available on Hugging Face under CC BY-NC 4.0 license.
What This Means
Mistral's entry into text-to-speech with a compact 4B parameter model and competitive pricing positions it against established players like ElevenLabs. The 70ms latency and 3-second voice cloning capability make it viable for real-time voice agent applications. The open-weight release under a non-commercial license follows Mistral's hybrid approach of offering both commercial API access and community model weights. Cross-lingual voice adaptation without explicit training is a notable capability that could simplify speech-to-speech translation pipelines.
Related Articles
Tencent Open-Sources AuK, a 1.5B-Parameter Speech Generation and Editing Model
Tencent has open-sourced AuK, a 1.5B-parameter foundation model for speech generation and editing that handles TTS, content editing, and audio enhancement through natural-language instructions. The release includes a distilled AuK-Flash variant for 4-step fast inference, both under MIT license.
Ex-OpenAI Researcher Launches Jev, an AI Model That Scores Options Instead of Generating Text
Startup TypeSafe AI has released Jev, a model built to score predefined answer options rather than generate text, claiming response times of 70 to 500 milliseconds. Co-founder Diogo Almeida, a former OpenAI researcher and InstructGPT co-author, says the model targets background classification tasks like sorting customer requests.
Anonymous Provider Launches Union Alpha, a Free 262K-Context Multimodal Model on OpenRouter
A third-party provider using the alias 'Stealth' has released Union Alpha on OpenRouter, a multimodal model with a 262K context window, currently free to use during its preview period. The model's developer remains anonymous, and OpenRouter states it is not the model's owner or operator.
TypeSafe Launches Jev, a Non-Generative 'Decision Model' Claiming Up to 200x Faster, 400x Cheaper Than Small LLMs
TypeSafe, founded by ChatGPT co-inventor Diogo Almeida, launched Jev on September 15, 2026 — a model that cannot generate free-form text but claims to classify, route, and score 20-200x faster and 40-400x cheaper than small frontier LLMs. Trained via a new method called RLCD, Jev targets production systems that use LLMs purely as structured judges or routers.
Comments
Loading...