Chatterbox Multilingual V3
0.5B parameter, Llama-based, multilingual, open-source TTS model.
Resemble AI released Chatterbox Multilingual V3 on June 10, 2026. It is an open-source text-to-speech (TTS) model based on Llama with 0.5 billion parameters. Just as GPT can understand and respond to any text, Chatterbox converts any text into natural-sounding speech in 25 languages. It can replicate a speaker's voice (zero-shot voice cloning) with just a short audio sample of 5-20 seconds, and its voice expressiveness can be precisely controlled, ranging from monotonous recitation to dramatic acting, through Classifier-Free Guidance (CFG) and exaggeration parameters.
Existing commercial TTS services offer high quality but have fundamental limitations, including cloud dependency, usage-based pricing, and data privacy restrictions. Chatterbox V3 breaks down these barriers with a fully open-source MIT license. In a blind evaluation, it achieved a 63.75% preference rate compared to ElevenLabs, while also enabling commercial deployment without royalties, revenue sharing, or usage restrictions. More noteworthy is the fact that Perceptual Threshold (PerTh) watermarking is built in by default. Utilizing psychoacoustic principles, it inserts a neural watermark that is inaudible to the human ear but survives MP3 compression, telephone codecs, and audio editing, preemptively meeting the AI-generated content labeling requirements of Article 50 of the EU AI Act, which will be implemented in August 2026.
V3 supports a total of 25 languages, including 21 base languages and 6 language packs (Mandarin Chinese, Hindi, Brazilian Portuguese, European Portuguese, Latin American Spanish, and European Spanish). It achieves production-level accuracy in major languages, with Italian CER (Character Error Rate) of 0.20%, German less than 0.20%, and English 0.65%. On a single H100 GPU, it achieves a TTFB (Time-to-First-Byte) of less than 300ms and an RTF (Real-Time Factor) of approximately 5x. When deployed with NVIDIA NIM, it offers a 2-39x improvement in throughput compared to unoptimized PyTorch. With 36,700 hours of training data, it improves upon V2 in terms of prompt deviation, repetition, accent drift, and speaker similarity issues. However, Korean (CER 70.90%) and Vietnamese (75.21%) still require additional work before they can be deployed in production.
💻 System Requirements
NVIDIA GPU with 4GB+ recommended (real-time generation with RTX 3060 or higher). CPU/Apple MPS compatible, but with reduced inference speed. TTFB under 300ms on H100, RTF ~5x
Approximately 0.5GB per model (0.5B parameters), around 4-5GB including the entire package and Language Pack
⚡ Installation
4-1. Quick Start
pip install chatterbox-tts
4-2. Basic Usage (Multilingual V3)
from chatterbox.mtl_tts import ChatterboxMultilingualTTS
model = ChatterboxMultilingualTTS.from_pretrained(device="cuda", t3_model="v3")
wav = model.generate(
"Bonjour, comment allez-vous aujourd'hui?",
language_id="fr",
audio_prompt_path="ref_speaker.wav"
)
4-3. Turbo Variant (Low-Latency English)
from chatterbox.tts_turbo import ChatterboxTurboTTS
model = ChatterboxTurboTTS.from_pretrained(device="cuda")
wav = model.generate(
"Hi there [chuckle], have you got a minute?",
audio_prompt_path="ref_clip.wav"
)
4-4. Source Installation
git clone https://github.com/resemble-ai/chatterbox.git
cd chatterbox
pip install -e .
🧬 Bio Use Cases
Building Multilingual Voice Agents
Integrate Chatterbox Multilingual V3 into a customer service chatbot to generate customized voice responses for key markets such as French, German, and Spanish. Deploy with NVIDIA NIM to handle hundreds of concurrent sessions, and ensure AI-generated voice traceability with PerTh watermarking.
Automating Audiobook and Podcast Production
Use a 5-20 second narrator sample for zero-shot voice cloning, then convert long-form text into natural-sounding speech. Fine-tune for content characteristics by setting the Exaggeration parameter to 0.7 or higher for dramatic expression, and the CFG weight to 0.3 for faster speaker pacing.
Producing Voice Content Compliant with the EU AI Act
In preparation for the enforcement of Article 50 in August 2026, automatically embed PerTh watermarks into all AI-generated voices. Maintain near 100% detection accuracy even after MP3 compression, telephone codec application, and editing. Provide the detection library on GitHub under a permissive license, enabling the creation of custom validation pipelines.
FAQ
What is Chatterbox Multilingual V3?
Resemble AI released Chatterbox Multilingual V3 on June 10, 2026. It is an open-source text-to-speech (TTS) model based on Llama with 0.5 billion parameters. Just as GPT can understand and respond to any text, Chatterbox converts any text into natural-sounding speech in 25 languages. It can replicate a speaker's voice (zero-shot voice cloning) with just a short audio sample of 5-20 seconds, and its voice expressiveness can be precisely controlled, ranging from monotonous recitation to dramatic acting, through Classifier-Free Guidance (CFG) and exaggeration parameters. Existing commercial TTS services offer high quality but have fundamental limitations, including cloud dependency, usage-based pricing, and data privacy restrictions. Chatterbox V3 breaks down these barriers with a fully open-source MIT license. In a blind evaluation, it achieved a 63.75% preference rate compared to ElevenLabs, while also enabling commercial deployment without royalties, revenue sharing, or usage restrictions. More noteworthy is the fact that Perceptual Threshold (PerTh) watermarking is built in by default. Utilizing psychoacoustic principles, it inserts a neural watermark that is inaudible to the human ear but survives MP3 compression, telephone codecs, and audio editing, preemptively meeting the AI-generated content labeling requirements of Article 50 of the EU AI Act, which will be implemented in August 2026. V3 supports a total of 25 languages, including 21 base languages and 6 language packs (Mandarin Chinese, Hindi, Brazilian Portuguese, European Portuguese, Latin American Spanish, and European Spanish). It achieves production-level accuracy in major languages, with Italian CER (Character Error Rate) of 0.20%, German less than 0.20%, and English 0.65%. On a single H100 GPU, it achieves a TTFB (Time-to-First-Byte) of less than 300ms and an RTF (Real-Time Factor) of approximately 5x. When deployed with NVIDIA NIM, it offers a 2-39x improvement in throughput compared to unoptimized PyTorch. With 36,700 hours of training data, it improves upon V2 in terms of prompt deviation, repetition, accent drift, and speaker similarity issues. However, Korean (CER 70.90%) and Vietnamese (75.21%) still require additional work before they can be deployed in production.
When should I use Chatterbox Multilingual V3?
0.5B parameter, Llama-based, multilingual, open-source TTS model.
What is a biomedical use case for Chatterbox Multilingual V3?
Building Multilingual Voice Agents: Integrate Chatterbox Multilingual V3 into a customer service chatbot to generate customized voice responses for key markets such as French, German, and Spanish. Deploy with NVIDIA NIM to handle hundreds of concurrent sessions, and ensure AI-generated voice traceability with PerTh watermarking.
📝 Update Notes
No update notes yet.
🧪 Related Code of Life
No related Code of Life posts yet.