chatterbox: Generate Speech from Text with Voice Cloning

Overview

Chatterbox is a family of open-source text-to-speech models from Resemble AI. It turns text into speech and supports voice cloning from a reference audio clip, with models aimed at general multilingual generation, low-latency English applications, and smaller CPU or on-device deployments.

It is a practical option for developers who want to run TTS from Python rather than rely exclusively on a hosted speech API. The repository includes model variants and usage examples, while the choice between them depends on language coverage, latency, and available compute.

Key Features

• Multilingual V3 supports 23 languages and voice cloning across languages.
• Turbo is a 350M-parameter English model designed for lower-compute, low-latency use.
• Nano is a 110M-parameter variant intended for CPU and resource-constrained deployments.
• Turbo and Nano support paralinguistic tags such as [laugh] and [chuckle].
• The original and multilingual models expose generation controls for speaking style and pacing.
• Generated audio includes Resemble AI's Perth watermark, with a documented method to check for it.
• Models can be loaded and used through Python APIs, with CUDA, CPU, and MPS device options shown in the examples.

Use Cases

• Voice-agent developers can use Turbo when they need English speech generation with voice cloning and expressive tags.
• Localization teams can try Multilingual V3 for speech generation and cross-language voice cloning across its supported languages.
• Developers targeting limited compute can evaluate Nano for CPU or on-device TTS.
• Creators can generate narration or expressive speech and adjust generation settings to tune delivery.

Getting Started

Install the package:

pip install chatterbox-tts

See the README for model selection, installation from source, and generation examples. The demo page provides sample audio.

Alternatives

• csm: CSM conditions speech generation on optional conversation context and requires CUDA-compatible hardware and Hugging Face checkpoints.
• MOSS-TTS: MOSS-TTS covers long-form narration, dialogue, sound effects, and streaming across multiple model architectures.
• VoxCPM: VoxCPM focuses on tokenizer-free multilingual synthesis and voice design, with VoxCPM2 requiring support for a 2B-parameter model.
• Spark-TTS: Spark-TTS targets bilingual synthesis and zero-shot voice cloning using a Qwen2.5-based model.

Considerations

• This repository provides models to run, not just a hosted API. Inference requires installing its Python dependencies and obtaining model weights; the README examples use CUDA, CPU, or MPS, but performance depends on the chosen model and hardware.
• Voice cloning requires a reference audio clip for the Turbo and Nano examples. Check the supported language list and choose a suitable model before building around a specific language or dialect.
• The README reports model sizes and performance claims, but those figures are project-provided and may not predict results on a particular device.
• Generated audio is watermarked. This can matter for workflows that need to detect or disclose synthetic speech.