Kitten-TTS-Mini-0.8-ONNX-WebGPU

KittenML/kitten-tts-mini-0.8 with its weights requantized 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 is published only with dynamic int8 quantization (MatMulInteger, ConvInteger, DynamicQuantizeLSTM), which WebGPU doesn't run. Its weights were dequantized back to float32, then quantized without calibration (round to nearest, symmetric, blocks of 32): matrix multiplications to 4-bit MatMulNBits, embedding tables to GatherBlockQuantized, convolution and LSTM weights to blocked int4 DequantizeLinear (opset 21). Activations stay float32. 78 MB → 51 MB.
• voices/<voice>.bin: the original voices.npz, one little-endian float32 [400, 256] file per voice.
• 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 a 4.5% word error rate (the original: 3.4%).

Use

Inputs and outputs as the original (input_ids [1, n], style [1, 256], speed [1] → waveform at 24 kHz): the ids of the chunk's eSpeak-style IPA between two 0s and the voice's style row at index min(characters in the chunk, 399). Drop the last 5000 samples of each chunk. See the original's kittentts package for the text preparation.

Licence

Apache 2.0, as the original (LICENSE).