Local, CPU-Friendly, High-Quality TTS (Text-to-Speech) with Kokoro
Points and comments are a snapshot, not live.
Kokoro TTS delivers high-quality, CPU-friendly speech synthesis with 82M parameters.
Kokoro, a model with only 82 million parameters, produces realistic speech in English, Mandarin, and Hindi, offering about 50 distinct voices. It can run entirely on CPU, reserving GPU for LLM inference. The Kokoro-FastAPI container image simplifies setup with Docker or Podman, serving an OpenAI-compatible API. Generation times on a short test paragraph: Intel Core i7-4770K (4.7 seconds), Apple M2 Pro (4.5 seconds), AMD Ryzen 7 8745HS (1.5 seconds). An alternative, Speaches, bundles Whisper STT but requires separate voice weight downloads.
What commenters are saying
Commenters broadly praise Kokoro's quality and ease of use, with several sharing practical integrations. Some note Pocket TTS as a superior alternative, especially for voice cloning, though opinions on PyTorch vs. ONNX performance differ. A common criticism is Kokoro's poor handling of single-word utterances, with a workaround involving cropping longer sentences. Users highlight successful deployments in accessibility products, home assistant, and intercom systems. Several share ports to iPhone ANE, CoreML, and browser-based implementations.