Kokoro

Kokoro is an 82M-parameter text-to-speech model that produces natural-sounding speech on a CPU. Despite the small size it handles English, Mandarin, and Hindi, and ships around 50 voices (mostly English). The weights and voices live on Hugging Face as hexgrad/Kokoro-82M; the hexgrad/kokoro repo is the Python inference library. Ariya Hidayat's how-to covers the fastest way to get it running.

Why it's interesting

It runs fast enough on old, GPU-less hardware to be practical. Hidayat's numbers for a short paragraph with the am_eric voice, best of three runs: a 12-year-old Intel Core i7-4770K did it in 4.7 seconds, an Apple M2 Pro in 4.5, and an AMD Ryzen 7 8745HS in 1.5. That leaves the GPU free for LLM inference — pair Kokoro on the CPU with a local model on the GPU and you can listen to answers instead of reading them. This is the same local-first argument titit-local-ai makes for LLMs, applied to speech, and the TTS counterpart to ghost-pepper's local Whisper speech-to-text.

Running it

The quickest path is the pre-built Kokoro-FastAPI container, which bundles the voice models (so the image is around 5 GB):

podman run -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-cpu

It serves a web UI at localhost:8880/web for a quick check, and — more usefully — an HTTP interface compatible with the OpenAI speech API, so existing OpenAI-TTS code works against it by pointing at the local base URL. Select a voice with the TTS_VOICE environment variable (e.g. am_eric); the full list is in the model's VOICES.md.

For a stack that also does speech-to-text, Speaches is another OpenAI-compatible container that adds Whisper STT alongside TTS. Unlike Kokoro-FastAPI it doesn't bundle voice weights — you download them through its API — but it's a single service for both directions.

Limitations

Short utterances are the weak spot. Asked to say a bare word like "six", Kokoro tends to pad it ("ah-six-ah"); the model's small size is the reason. The practical workaround is to feed it a full sentence ("The word is: six"), then crop the audio using the per-word timestamps the API returns — reliable, if a little flat in intonation. Users also report the male voices lag the better female ones in quality, and SSML / fine inflection control is limited. For homographs it accepts inline IPA pronunciation hints, which covers the common mispronunciation cases.

Kokoro's language coverage is fixed; for languages it doesn't support (Dutch, for instance) you're back to larger, slower models. People wanting voice cloning or finer emotional control tend to reach for other small models — Pocket TTS, Supertonic, Chatterbox — but for plain, fast, private narration Kokoro is the common default, and it has been ported widely (WASM/WebGPU in the browser, CoreML and the Apple Neural Engine, ONNX).

8,030 stars, Apache-2.0 license. Repo: https://github.com/hexgrad/kokoro