Paradee-8M-ONNX-WebGPU

sahilmahendrakar/Paradee-8M-v1.1 with its weights quantized for running in a web browser with ONNX Runtime Web's WebGPU backend, as used by Quick Inference.

Changes from the original

• onnx/model.onnx: the original float32 graph quantized without calibration (round to nearest, symmetric, blocks of 32) to 8 bits: matrix multiplications to 8-bit MatMulNBits, convolution and LSTM weights to blocked int8 DequantizeLinear (opset 21). Activations stay float32. 35 MB → 15 MB. (4 bits made it audibly buzz; the original's own int8 file uses dynamic quantization, which WebGPU doesn't run.)
• A small float64 section (the pitch-phase accumulator) computes in float32, as WebGPU has no float64.
• config.json: the original.

These are modified files: their outputs are close to, but not identical to, the original model's. On eight test sentences, Whisper small.en heard it with an 8.0% word error rate, the same as the original.

Use

As the original's paradee package: Misaki phonemes, their ids from the config's vocab between two 0s (at most 510), and speed [1] → waveform at 24 kHz, one sentence at a time.

Licence

Apache 2.0, as the original (LICENSE).