Known Bug: Audio output is silent (all NaN samples)
Kokoro TTS in WASM produces NaN audio data during ONNX inference. This is a
known upstream issue (#2236)
in sherpa-onnx — the ONNX runtime on WASM/CPU generates NaN values in the model output tensor.
Even building from source does not fix it. The model loads and tokenizes text correctly, but inference output is all NaN.
Waiting for an upstream fix from the sherpa-onnx team.
Loading WASM module...
Note: Kokoro runs on a single CPU thread in WASM — generation may be slow for longer text.