Frequently Asked Questions

Voxtral TTS is a text-to-speech model created by Mistral AI, released March 2026. It converts text into natural, emotionally expressive speech with support for voice cloning. We provide a free web interface so anyone can use it without API keys or coding.

No. You can generate speech immediately without signing up. Creating an account (via Google) unlocks higher usage limits and lets you save preferences.

Each generation supports up to 2,000 characters. For longer content like articles or book chapters, split your text into sections and generate them individually. The model handles paragraph transitions and punctuation-aware pacing automatically.

Yes. Every user gets free credits daily — no credit card required. Free credits are sufficient for most personal and small-project use. Signed-in users receive additional credits.

Upload a 2–25 second audio clip of any voice. Voxtral analyzes the speaker's accent, pitch, rhythm, and emotional texture in a single forward pass — no fine-tuning or training required. The cloned voice can then speak any text you provide.

MP3, WAV, and WEBM files are supported. For best results, use 5–10 seconds of clear conversational speech with minimal background noise. Studio-polished recordings actually produce less natural clones than casual speech.

Nine languages: English, French, Spanish, Portuguese, Italian, Dutch, German, Hindi, and Arabic. All languages support both preset voices and voice cloning.

Yes — this is called cross-lingual voice cloning. Upload a French speaker's voice and generate fluent German speech while preserving the original speaker's identity and accent characteristics. The model handles language-specific phonetics automatically.

Yes. Voxtral uses a "voice-as-an-instruction" approach. It follows the intonation, rhythm, and emotional rendering of the reference voice without needing prosody or emotion tags. Write excited text and the output sounds excited. Write a question and intonation rises naturally.

Voxtral TTS achieves ~90ms model processing latency. End-to-end time-to-first-audio is approximately 0.8 seconds for PCM streaming format and about 3 seconds for MP3 files. This makes it suitable for real-time voice agent applications.

MP3 for general use (podcasts, web embedding, social media), WAV for lossless quality (video production, professional editing), and PCM for real-time streaming applications.

Just 4 billion parameters — compact enough to run on laptops, smartphones, and edge devices with as little as 3GB of RAM. The open-weight model is available on Hugging Face under a permissive license for self-hosting.

Yes. The model weights are open-source. At 4B parameters and 3GB RAM, you can deploy on consumer hardware for full data sovereignty — no audio ever leaves your infrastructure.

Voxtral TTS is the only open-weight model in its quality tier. It matches or exceeds ElevenLabs and OpenAI TTS on naturalness benchmarks while offering advantages in latency (90ms), self-hosting capability, and cost (free credits, $0–$16/M characters via API).

No. Your text inputs and uploaded voice references are never used to train or fine-tune AI models. All data processing is for your generation only.

Uploaded voice references are deleted within 1 hour. Generated audio is available during your session for download. We do not permanently archive any user content.

Voice cloning analyzes acoustic properties of audio you upload (pitch, rhythm, timbre). This data is processed transiently for generation and is not stored, shared, or used beyond your immediate request.

Email us at contact@voxtraltts.net. We respond within 5 business days. For GDPR or CCPA rights requests, we process them within 30 days.