pocket-tts vs voicebox
Kyutai's 100M-parameter CPU-only TTS — streaming audio in ~200ms, ~6× real-time on two laptop cores, voice cloning, six languages. pip install and it talks. — versus — Local-first AI voice studio: clone voices, generate speech via 7 TTS engines in 23 languages, dictate system-wide with Whisper, and give any MCP-aware agent a cloned voice. Tauri, MLX/CUDA.
Same say-it-locally job, different shape: pocket-tts is a pip-installable 100M CPU library for embedding speech in your own code; Voicebox is a full desktop studio with cloning, dictation, effects and MCP agent integration.
| pocket-tts | voicebox | |
|---|---|---|
| Stars | 7.6k | 42k |
| Forks | 776 | 5.0k |
| Language | Python | TypeScript |
| License | MIT | MIT |
| Last activity | 23 days ago | 3 days ago |
| Topics | local, voice | voice |
| Curated connections | 3 | 2 |
pocket-tts — the curator's take
This is the TTS you add when the rest of your stack is already local: no GPU, no API key, ~200ms to first audio on two CPU cores, and a port ecosystem (WASM, MLX, ONNX, C++, Home Assistant) that means it runs basically anywhere — the natural voice for an Ollama-powered assistant. When NOT: you need studio-grade expressiveness or high-concurrency serving — it's a batch-of-1, small-model design, not a production voice API; and the polished voice catalog is English-heaviest. Voice cloning needs 20s of audio — use it with consent.
voicebox — the curator's take
The whole voice I/O loop in one local app — ElevenLabs (output) plus WisprFlow (input) territory without the cloud. Zero-shot cloning, 50+ preset voices, paralinguistic tags via Chatterbox Turbo, a stories/timeline editor, and a global dictation hotkey with local-LLM cleanup. The agent hook is the standout: one `voicebox.speak` MCP tool call and Claude Code or Cursor talks in a voice you cloned, with per-agent voice binding and an always-visible on-screen pill so no agent speaks silently; a REST /speak covers non-MCP harnesses. NOT a library — it's a desktop product (Tauri): Linux is build-from-source, models are hefty downloads, and engine quality varies (only Turbo interprets [laugh]-style tags; others read them literally).