🪴 Brain

Pocket TTS

3 min read

Pocket TTS

🚀 Description

kyutai-labs/pocket-tts — a text-to-speech model that fits in your CPU (and pocket). Kyutai’s answer to the “TTS needs a GPU or a paid API” assumption: a 100M-parameter model that generates speech ~6× faster than real-time on a MacBook Air M4 using only 2 CPU cores, with ~200ms latency to the first audio chunk.

Generating audio is pip install pocket-tts (or uvx pocket-tts generate) and one function call. Kyutai explicitly tried GPU execution and found no speedup over CPU — at batch size 1 with a model this small, the GPU buys nothing.

🧩 Features

  • CPU-only by design — 100M params, 2 cores, PyTorch 2.5+ (CPU build), Python 3.10–3.14.
  • Streaming + low latency — ~200ms to first chunk; handles infinitely long text inputs.
  • Voice cloning — pass any wav file as --voice; export-voice converts audio → fast-loading safetensors embedding (just a KV-cache read). Curated voice catalog on kyutai/tts-voices.
  • 6 languages — English, French, German, Portuguese, Italian, Spanish; non-English also in higher-quality 24-layer variants (--language italian_24l). No Polish.
  • Three interfaces — CLI (generate / serve / export-voice), local HTTP server with web UI, and a plain Python API (TTSModel.load_model() → generate_audio()).
  • Runs in the browser — small enough for WASM; community ports: Rust/XN, ONNX Runtime Web, Candle, jax-js.
  • Big ecosystem — MLX (Apple Silicon), C++ single-file runtime (PocketTTS.cpp), sherpa-onnx (12 language bindings, Raspberry Pi/Jetson), C#/.NET port, OpenAI-compatible streaming servers, Home Assistant (Wyoming), ComfyUI node, Discord bots, audiobook tools.

Reasoning for

Fills the free local tier of my TTS stack thinking — the same slot Chatterbox TTS holds in Travelcast AI’s fallback chain, but radically cheaper to run: no GPU, no torch-CUDA install, ~instant startup via uvx. For agent voice output (Voicebox pattern) or any pipeline where ElevenLabs-quality is overkill, a 100M CPU model with streaming and cloning is the new floor. Main limitation for my use: no Polish — see TTS Engines Comparison (Polish), where Azure still wins Polish quality. The prohibited-use policy requires lawful consent for voice cloning.

Alternatives considered

  • Chatterbox TTS — open-source local TTS with cloning, but torch-heavy; Pocket TTS is far lighter.
  • Voicebox — local-first voice studio (7 TTS engines, STT, MCP) — an app, not a library; could adopt Pocket TTS as an engine.
  • ElevenLabs — best-in-class quality and languages (paid API); Pocket TTS is the zero-cost offline end of that spectrum.
  • OpenAI TTS — cloud API voices; no offline story.

🔗 Links

🔗 Related notes

Template: tool

Graph View

Backlinks