An ONNX export of Peacockery/hubert-base-phoneme-en
(HuBERT-base fine-tuned with CTC on English ARPAbet phones), with its MatMul/Gemm weights dynamically quantised to
int8. One file: phoneme.onnx (~120 MB). It runs on the phone, CPU only, through onnxruntime-react-native in the
Lingubi app, which scores how well a child's attempt fits a target word.
Peacockery/hubert-base-phoneme-en, Apache-2.0.torch.onnx.export (opset 17) of the model with log_softmax added on its logits, then
onnxruntime.quantization.quantize_dynamic(weight_type=QInt8, op_types_to_quantize=["MatMul", "Gemm"]).
audio: float32 [1, n], 16 kHz mono, normalised by the caller to zero mean / unit variance.logprobs: float32 [1, frames, 43], log-probabilities, one frame per 20 ms. The vocabulary names 41
tokens; the last two columns are never trained. Read the width from the output, do not hard-code 41.[PAD], id 40.