Kokoro
Open-weight text-to-speech model with 82 million parameters, released under Apache 2.0. It runs on your own machine through a Python package, covers 8 languages including French, and its weights are only a 327 MB download.
Strengths
- Apache 2.0 weights, commercial use included, no clause to negotiate and no API key to manage
- Just 82 million parameters and 327 MB on disk: the lightest of its class, inference stays fast without exotic hardware
- Covers 8 languages (English, French, Spanish, Italian, Brazilian Portuguese, Hindi, Japanese, Mandarin) with a voice bank included
- Nothing leaves your machine: the text you have it read never transits through any provider
Limitations
- No voice cloning at all: you are locked into the shipped voice bank, you cannot use your own voice or a brand voice
- No emotion control and no intonation direction: on a long text the delivery flattens out and it shows
- A system dependency on espeak-ng sits next to the Python package: on a server that is one more component to deploy and maintain
- You host inference yourself: a machine kept running, a Python environment to keep current, and noticeable latency without a GPU
Best for
- Audio read-aloud of articles or documentation, where volume matters more than vocal charm
- Voice agent prototypes, to validate a flow without opening an account or burning quota
- Contexts where the spoken text is confidential and cannot be sent to a third party