The Canadian startup Resemble AI recently introduced Chatterbox, its first open-source TTS (Text-to-Speech) model. Distributed under the MIT license, this voice cloning model positions itself as a credible alternative to proprietary market solutions, while introducing unprecedented features for an open-source model.
Chatterbox is based on a 0.5 billion parameter architecture, trained on 500,000 hours of cleaned data.
Key model features:
- Zero-Shot Voice Cloning: With just a few seconds of reference audio, Chatterbox can mimic any voice without requiring additional training;
- Emotion Control: Unlike other speech synthesis models, Chatterbox allows the adjustment of the emotional intensity of speech, ranging from a monotone to dramatic expressiveness, according to user needs;
- Real-Time Speech Synthesis: Thanks to alignment-based generation, the model operates faster than real-time inference, making it ideal for voice assistants, video games, and interactive applications.
- Security Watermark: Every generated audio file includes a perceptual watermark (PerTh Watermarker), ensuring transparency and traceability of the generated content.
The use of Chatterbox is simplified thanks to a dedicated Python library (
chatterbox-tts), compatible with CUDA. The model can be initialized locally or from pre-trained models. Developers can also provide custom voice samples (audio prompts) to adjust style or target voice.Resemble AI compared Chatterbox to proprietary market models.
Chatterbox vs Competition
| Feature | Chatterbox | ElevenLabs | Google TTS | Azure TTS |
|---|---|---|---|---|
| License | MIT (Free) | Proprietary | Proprietary | Proprietary |
| Emotion Control |  Advanced | Basic |  | |
| Latency | <200 ms | ~300 ms | ~400 ms | ~500 ms |
| User Preference | 63.75% | 36.25% | N/A | N/A |
| Watermarking | Integrated | | | |
| Voice Cloning | Yes | Yes | | Limited |
In a comparative test conducted by Podonos, listeners preferred Chatterbox in 63.75% of cases over the proprietary model from ElevenLabs, which is considered one of the market leaders.
Resemble AI provides a demonstration interface via Hugging Face (Gradio), allowing users to test the model without local installation. For more intensive or critical uses, the company offers a commercial version of the TTS engine with latency below 200 ms.