Voxtral Tts API

Voxtral TTS 🎙️

Text-to-speech via Mistral's Voxtral TTS API. 4B parameter model, 70ms latency, $0.016/1k chars.

Setup

export MISTRAL_API_KEY="your-key"
# or
cat ~/.openclaw/secrets/mistral.key

Quick Start

# Generate speech with a preset voice
curl -s "https://api.mistral.ai/v1/audio/speech" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Hello, this is a test.",
    "voice_id": "<voice-id>",
    "response_format": "mp3"
  }' | jq -r '.audio_data' | base64 -d > output.mp3

# Clone a voice from a 3-5 second sample
REF_AUDIO=$(base64 -w0 reference.wav)
curl -s "https://api.mistral.ai/v1/audio/speech" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"input\": \"Speaking with a cloned voice.\",
    \"ref_audio\": \"$REF_AUDIO\",
    \"response_format\": \"mp3\"
  }" | jq -r '.audio_data' | base64 -d > cloned.mp3

API Reference

Generate Speech

POST https://api.mistral.ai/v1/audio/speech

Auth: Authorization: Bearer $MISTRAL_API_KEY

Response (non-streaming):

{"audio_data": "<base64-encoded-audio>"}

Response (streaming): SSE events:

event: speech.audio.delta
data: {"audio_data": "<base64-chunk>", "type": "speech.audio.delta"}

event: speech.audio.done
data: {"type": "speech.audio.done", "usage": {"prompt_tokens": 42, ...}}

Note: Either voice_id or ref_audio should be provided. If neither is given, a default voice is used.

Voice Management

List voices:

curl -s "https://api.mistral.ai/v1/audio/voices?limit=50" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" | jq '.items[] | {id, name, gender, languages}'

Create a saved voice (from audio sample):

SAMPLE=$(base64 -w0 my-voice.wav)
curl -s -X POST "https://api.mistral.ai/v1/audio/voices" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"name\": \"My Custom Voice\",
    \"sample_audio\": \"$SAMPLE\",
    \"sample_filename\": \"my-voice.wav\",
    \"gender\": \"male\",
    \"languages\": [\"en\", \"de\"]
  }" | jq '{id, name}'

Update voice metadata:

curl -s -X PATCH "https://api.mistral.ai/v1/audio/voices/<voice-id>" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"name": "Updated Name", "tags": ["narrator", "warm"]}'

Delete voice:

curl -s -X DELETE "https://api.mistral.ai/v1/audio/voices/<voice-id>" \
  -H "Authorization: Bearer $MISTRAL_API_KEY"

Languages

9 languages supported natively:

Features

Voice Cloning

Pass ref_audio (base64, 3-25 seconds) for zero-shot voice cloning. Captures accent, rhythm, intonation, pauses, and emotional patterns. No pre-registration needed.

Cross-Lingual Synthesis

The model handles cross-lingual voice adaptation: a French voice prompt with German text produces natural German speech with a French accent. Useful for translation pipelines.

Streaming

Set stream: true for SSE streaming. Useful for real-time voice agents. Events: speech.audio.delta (audio chunks) and speech.audio.done (completion + usage).

Long Text

The model natively generates up to 2 minutes of audio. The API handles longer text automatically with smart interleaving.

Pricing

$0.016 per 1,000 characters. A typical 500-character message costs ~$0.008.

Use Cases

Shell Helper

# One-liner: text to mp3 file
voxtral() {
  curl -s "https://api.mistral.ai/v1/audio/speech" \
    -H "Authorization: Bearer $MISTRAL_API_KEY" \
    -H "Content-Type: application/json" \
    -d "{\"input\": \"$1\", \"voice_id\": \"${2:-}\", \"response_format\": \"mp3\"}" \
    | jq -r '.audio_data' | base64 -d > "${3:-output.mp3}"
  echo "Saved: ${3:-output.mp3}"
}

# Usage:
# voxtral "Hallo Welt!" "voice-id" "hallo.mp3"
# voxtral "Quick test"  # uses default voice, saves to output.mp3

Comparison

v1.0 -- Voxtral TTS API skill. ManniTheRaccoon 2026-03-30. 🎙️