Pocket TTS — Lightweight Text-to-Speech for CPU

Introduction

Pocket TTS is a lightweight text-to-speech model developed by Kyutai Labs that generates natural-sounding speech using only your CPU. It fits comfortably in memory on modest hardware, making it ideal for edge devices, CI pipelines, and local development environments where GPU access is unavailable.

What Pocket TTS Does

• Generates natural multi-speaker speech from text on CPU
• Supports multiple languages with a single compact model
• Produces audio at faster-than-real-time speeds on modern CPUs
• Offers voice cloning from short reference clips
• Provides streaming synthesis for real-time applications

Architecture Overview

Pocket TTS uses a distilled transformer architecture optimized for CPU inference. The model compresses speech representations into discrete tokens, then decodes them through a lightweight vocoder. Quantization and operator fusion keep the total model size small while maintaining audio quality comparable to larger GPU-based systems.

Self-Hosting & Configuration

• Install via pip with no external dependencies beyond PyTorch CPU
• Model weights download automatically on first run (~200MB)
• Configure voice, speed, and language via simple API parameters
• REST API wrapper available for microservice deployment
• Docker image included for containerized environments

Key Features

• Sub-200MB total model size fits in constrained environments
• No GPU or CUDA installation required
• Real-time factor below 0.5x on modern laptop CPUs
• Permissive license for commercial use
• Clean Python API with async support

Comparison with Similar Tools

• Coqui TTS — larger models requiring GPU for real-time; Pocket TTS is CPU-native
• Piper — fast but limited voice variety; Pocket TTS supports cloning
• Bark — high quality but very slow on CPU; Pocket TTS optimized for speed
• Kokoro — similar size but fewer languages supported

FAQ

Q: What hardware is needed? A: Any x86-64 or ARM64 CPU with 1GB free RAM. No GPU needed.

Q: Which languages are supported? A: English, French, German, Spanish, and Mandarin at launch with community additions.

Q: Can I fine-tune on my own voice? A: Yes. A fine-tuning script is provided requiring only 30 seconds of reference audio.

Q: What audio formats are output? A: WAV by default, with optional MP3 and OGG export.

Sources

• https://github.com/kyutai-labs/pocket-tts