ONNX conversion of Peacockery/hubert-base-phoneme-en — a HuBERT-base (95M) fine-tune on LibriSpeech-960 for ARPAbet phoneme CTC (base: facebook/hubert-base-ls960).
The phoneme-CTC acoustic, updated October 2026 to a speaker-disjoint fine-tune of the source model on human L2 English. The previous export remains available in this repository's history.
ConvInteger kernel.
The int8 export was validated against fp32 on held-out utterances.[UNK]/[PAD]/<s>/</s>; CTC blank = [PAD] (id 40).
Input: zero-mean/unit-variance 16 kHz mono float samples.87d0debefeef6ebbc2b5e5bcdafeeebf171714c424abe517c66cedac198eba06The original-weight int8 export of the source model with the pre-head hidden
state exposed — one added Identity node and two extra graph outputs:
hidden_states (the layer-11 final-layer-norm tensor the CTC head consumes,
768 float values per 20 ms frame) and hidden_mid (the layer-6 equivalent).
Its logits match the previous export of hubert_phoneme_int8.onnx on real
takes; it is unchanged by the October 2026 fine-tune of the acoustic file.
de142bc9db527a97ec61d2e8b42722acb78d20c28ee046d2a3e0d95cfbb49f74hubert_phoneme_emb.onnx 122 MBhubert_phoneme_int8.onnx 122 MB