Voxtral-4B-TTS-2603

Voxtral TTS is a frontier, open-weights text-to-speech model that’s fast, instantly adaptable, and produces lifelike speech for voice agents. The model is released with BF16 weights and a set of reference voices. These voices are licensed under CC BY-NC 4, which is the license that the model inherits.

For more details, see our:

• 🔊 Demo
• ✍️ Blog post
• 🔬 Research Paper

Key Features

Voxtral TTS delivers enterprise-grade text-to-speech for production voice agents, with the following capabilities:

• Realistic, expressive speech with natural prosody and emotional range across 9 major languages, with support for diverse dialects
• Text-to-Speech generation with 20 preset voices and easy adaptation to new voices
• Multilingual support: English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic, and Hindi
• Very low latency with fast time-to-first-audio, plus streaming and batch inference support
• 24 kHz audio output in WAV, PCM, FLAC, MP3, AAC, and Opus formats
• Production-ready performance for high-throughput, real-time voice agent workflows

[!Tip] For voice customization, visit our AI Studio.

Use Cases

• Customer support and call center infrastructure.
• Financial services. -- with video demo on banking KYC voice agents.
• Manufacturing and industrial operations.
• Public services and government.
• Compliance and risk.
• Supply chain and logistics.
• Automotive and in-vehicle systems.
• Sales and marketing.
• Real-time translation.

[!Warning] Responsible Use - You are responsible for complying with applicable laws and avoiding misuse.

Benchmark Results

• Measured using vllm_omni/examples/offline_inference/voxtral_tts/end2end.py.
• Input: 500-character text with a 10-second audio reference.
• Hardware: single NVIDIA H200.
• vllm version: v0.18.0.

Note: The RTF in end2end.py uses an inverted formula (higher = better). The table below converts it back to the standard RTF convention (lower = better)

Concurrency	Latency	RTF	Throughput (char/s/GPU)
1	70 ms	0.103	119.14
16	331 ms	0.237	879.11
32	552 ms	0.302	1430.78

Usage

The model can also be deployed with the following libraries:

• vllm-omni (recommended): See here

vLLM Omni (recommended)

[!Tip] We've worked hand-in-hand with the vLLM-Omni team to have production-grade support for Voxtral 4B TTS 2603 with vLLM-Omni. Special thanks goes out to Han Gao, Hongsheng Liu, Roger Wang, and Yueqian Lin from the vLLM-Omni team.

Installation

Make sure to install vllm from the latest (>= 0.18.0) pypi package. See here for a full installation guide.

uv pip install -U vllm

Next, you should install vllm-omni with vllm-omni >= 0.18.0.

uv pip install vllm-omni --upgrade  # make sure to have >= 0.18.0

Alternatively, you can also make use of a ready-to-go docker image on the docker hub.

Installing vllm >= 0.18.0 should automatically install mistral_common >= 1.10.0 which you can verify by running:

python3 -c "import mistral_common; print(mistral_common.__version__)" # should print >= 1.10.0

Serve

Due to size and the BF16 format of the weights - Voxtral-4B-TTS-2603 can run on a single GPU with >= 16GB memory.

vllm serve mistralai/Voxtral-4B-TTS-2603 --omni

Client

import io
import httpx
import soundfile as sf
 
BASE_URL = "http://<your-server-url>:8000/v1"
 
payload = {
    "input": "Paris is a beautiful city!",
    "model": "mistralai/Voxtral-4B-TTS-2603",
    "response_format": "wav",
    "voice": "casual_male",
}
 
response = httpx.post(f"{BASE_URL}/audio/speech", json=payload, timeout=120.0)
response.raise_for_status()
 
audio_array, sr = sf.read(io.BytesIO(response.content), dtype="float32")
print(f"Got audio: {len(audio_array)} samples at {sr} Hz")

# you can play the audio with a library like `sounddevice.play` for example

Demo

To run it:

git clone https://github.com/vllm-project/vllm-omni.git && \
cd vllm-omni && \
uv pip install gradio==5.50 && \
python examples/online_serving/voxtral_tts/gradio_demo.py \
  --host <your-server-url> \
  --port 8000

Alternatively you can also try it out live here ➡️ HF Space.

License

The provided voice-references compatible with this model are licensed under CC BY-NC 4, e.g. from EARS, CML-TTS, IndicVoices-R and Arabic Natural Audio datasets. Thus, this model inherits the same license.

You must not use this model in a manner that infringes, misappropriates, or otherwise violates any third party’s rights, including intellectual property rights.