chatterbox: Generate Speech with Resemble AI TTS Models

Overview

Chatterbox is Resemble AI’s family of open-source text-to-speech models. It turns text into speech and supports voice cloning from a reference audio clip. The model range includes multilingual speech generation, English-focused models for lower-latency use, and a smaller model intended for CPU and on-device inference.

The Python package can run on CUDA, CPU, or Apple’s MPS backend. It is aimed at developers who want to generate speech in code, including teams building voice agents, narration tools, or multilingual applications. The repository also describes built-in Perth audio watermarking for generated audio.

Key Features

• Generate English speech with the original Chatterbox model or Chatterbox-Turbo.
• Clone a voice by providing a reference audio clip.
• Generate multilingual speech with Chatterbox Multilingual V3, which supports 23 listed languages.
• Use dedicated single-language model fine-tunes for selected languages and regional variants.
• Add paralinguistic tags such as [laugh] and [chuckle] with Turbo and Nano.
• Run the smaller Nano model on CPU, or use CUDA and MPS where available.
• Embed and extract Perth perceptual-threshold watermarks in generated audio.

Use Cases

• Voice-agent developers can use Turbo for English speech generation with voice prompts and expressive tags.
• Localization teams can generate speech across the supported languages, or try a dedicated language fine-tune where available.
• Creators can produce narration and expressive audio with voice cloning and generation controls.
• Teams with tight compute or memory budgets can evaluate Nano for CPU or on-device inference.

Getting Started

Install from PyPI:

pip install chatterbox-tts

See the README for model-specific examples and source installation instructions.

Alternatives

• chatterbox-vllm: A vLLM deployment of Chatterbox focused on batched GPU inference, rather than the original model package and its broader usage options.
• MOSS-TTS: MOSS-TTS adds sound generation, multi-speaker dialogue, and real-time streaming, while Chatterbox focuses on speech synthesis and voice cloning.
• Spark-TTS: Spark-TTS uses Qwen2.5 and targets controllable Chinese and English speech, while Chatterbox offers multilingual models and smaller English variants.
• GPT-SoVITS: GPT-SoVITS combines zero-shot voice cloning with a WebUI and speaker fine-tuning, while Chatterbox provides pretrained synthesis models.

Considerations

The project requires Python 3.10 or newer; the README says development and testing used Python 3.11 on Debian 11. The examples use PyTorch audio tooling, and the README presents CUDA as an option rather than a requirement: CPU and MPS are also shown, while Nano is specifically intended for CPU and on-device use. Check the model’s resource needs against your deployment environment before choosing a variant. The repository is marked quiet by its recent activity data, with no commits in the last 90 days and no release in over a year, so teams should assess their tolerance for a slower-moving upstream project.