voxtral-tts

name

voxtral-tts

description

Generate speech from text using Mistral's Voxtral TTS model locally on Apple Silicon via mlx-audio. 20 preset voices across 9 languages (English, French, German, Spanish, Portuguese, Italian, Dutch, Arabic, Hindi). Does NOT support Chinese/Japanese/Korean. Use when: user asks to "generate speech", "text to speech", "TTS", "voxtral", "read aloud", "speak this", or needs high-quality multilingual voice synthesis locally. Preferred over cloud TTS for privacy and cost.

Voxtral TTS

Local text-to-speech on Apple Silicon using Mistral's Voxtral-4B via MLX.

Requirements

• Apple Silicon Mac (M1+)
• Python 3.10+
• ~2.5GB RAM (4-bit quantized) or ~8GB (bf16)

Setup

pip install "mlx-audio @ git+https://github.com/lucasnewman/mlx-audio.git@main"

Note: As of March 2026, Voxtral support requires mlx-audio from git (PR #607). Once released to PyPI, pip install -U mlx-audio will suffice.

Quick Usage

from mlx_audio.tts.utils import load
import numpy as np
import soundfile as sf

model = load("mlx-community/Voxtral-4B-TTS-2603-mlx-4bit")

for result in model.generate(text="Hello world!", voice="cheerful_female"):
    audio = np.array(result.audio)
    sf.write("output.wav", audio, result.sample_rate)  # 24000 Hz

Always use result.sample_rate — never hardcode the sample rate.

Model Variants

Model	Size	RAM	Speed
mlx-community/Voxtral-4B-TTS-2603-mlx-4bit	~2.5GB	~3GB	RTF ~0.70
mlx-community/Voxtral-4B-TTS-2603-mlx-6bit	~3.5GB	~4GB	RTF ~0.75
mlx-community/Voxtral-4B-TTS-2603-mlx-bf16	~8GB	~9GB	Best quality

4-bit recommended for most use cases. RTF < 1.0 means faster than real-time.

Voices (20 total)

English (5)

casual_female, casual_male, cheerful_female, neutral_female, neutral_male

Other Languages (15)

Language	Voices
Portuguese 🇧🇷	pt_male, pt_female
Dutch 🇳🇱	nl_male, nl_female
Italian 🇮🇹	it_male, it_female
French 🇫🇷	fr_male, fr_female
Spanish 🇪🇸	es_male, es_female
German 🇩🇪	de_male, de_female
Arabic 🇸🇦	ar_male
Hindi 🇮🇳	hi_male, hi_female

Script Usage

Generate speech from the command line:

python scripts/voxtral_generate.py --text "Hello world" --voice cheerful_female --output output.wav

Options:

• --text — Text to synthesize (required)
• --voice — Voice name (default: cheerful_female)
• --model — Model variant (default: mlx-community/Voxtral-4B-TTS-2603-mlx-4bit)
• --output — Output file path (default: output.wav)
• --format — Output format: wav, ogg, mp3 (default: wav; ogg/mp3 require ffmpeg)

First Run

The first run downloads the model from HuggingFace (~2.5GB for 4-bit). The script auto-detects this and prints a notice. Subsequent runs load from cache in ~2-5 seconds.

Performance (M5 MacBook Pro 32GB, 4-bit)

• Load time: ~2-5s (first load downloads model)
• Generation RTF: ~0.70 (10s audio ≈ 7s generation)
• Sample rate: 24000 Hz
• Languages not supported: Chinese, Japanese, Korean — use spark-tts skill for Chinese

Integration Pattern

from mlx_audio.tts.utils import load
import numpy as np
import sounddevice as sd

model = load("mlx-community/Voxtral-4B-TTS-2603-mlx-4bit")

def speak(text: str, voice: str = "cheerful_female"):
    for result in model.generate(text=text, voice=voice):
        audio = np.array(result.audio)
        sd.play(audio, result.sample_rate)
        sd.wait()

Telegram Voice Message

To send as Telegram voice, convert to OGG Opus:

ffmpeg -y -i output.wav -c:a libopus -b:a 96k output.ogg

Then send with asVoice: true.