Kokoro-82M
hexgrad/Kokoro-82M · Apache-2.0
Studio-quality TTS from an 82M model you can run on a CPU.
Kokoro packs a surprising amount of naturalness into 82 million parameters, with multiple voices and languages — small enough to ship in a browser or on edge hardware.
Score
Setup at a glance
pip install kokoro Deep dive
Kokoro-82M is an 82-million-parameter text-to-speech model with quality far above what its size suggests — it runs comfortably on CPU, which makes it the default pick when TTS needs to be cheap, local or on-device.
The model card ships distinct voices (EVA, SAM and others) with sample audio, and the pip package (kokoro) plus soundfile and espeak-ng is the whole install.
Independent benchmarks track it on Artificial Analysis' text-to-speech family page, and hosted inference is available through DeepInfra if you would rather call an API than run it.