Kyutai Labs released Pocket TTS, a text-to-speech model that runs on CPUs. It uses 100 million parameters and needs only two CPU cores. It achieves around 200 milliseconds latency for the first audio chunk. It runs about six times faster than real-time on a MacBook Air M4. It supports English, French, German, Portuguese, Italian, Spanish and Dutch. It offers voice cloning from a WAV file. It can process infinitely long text inputs. It provides a Python API, a CLI and a local HTTP server. It can also run in a browser on the client side. Training code is now available for community use.
What changed
TTS models typically required GPUs or cloud APIs to run efficiently.
What it unlocks
Running a lightweight text-to-speech model locally on a CPU with low latency and minimal resource use.
- 100M parameters
- ~200ms first audio chunk latency
- ~6x real-time on MacBook Air M4 CPU
- 2 CPU cores
What you need to act on it
- Python 3.10-3.14
- PyTorch 2.5+ (CPU-only)
Sources