Kokoro-82M Inference Endpoints handler
Serves hexgrad/Kokoro-82M on a
dedicated Hugging Face Inference Endpoint via a custom handler.py.
Request
{"inputs": "Text to speak.", "parameters": {"voice": "af_heart", "speed": 1.0}}
voice is a Kokoro voice id (af_*/am_* are American female/male).
Response
{"audio_b64": "<base64 WAV, 24 kHz mono s16>", "sample_rate": 24000}
Deploy: the model is 82M params — an Nvidia L4 is ample; CPU works too.
First boot installs requirements.txt and downloads the ~330 MB model.