fuann/hmamba-full-distilled

Distilled HMamba for SpeechOcean762

This repository contains the HMamba inference core used for phone-, word-, and utterance-level English pronunciation assessment. It is intended to be used with the run_hmamba_infer.sh pipeline from the HMamba repository.

This is not a waveform-to-score model by itself. The script supplies the required Kaldi GOP, phone alignment, Wav2Vec2, HuBERT, WavLM, and raw acoustic features.

bash run_hmamba_infer.sh \
  --model-source USERNAME/hmamba-full-distilled \
  --wav example.wav \
  --utt-id example_0001 \
  --text "REFERENCE TRANSCRIPT"

Files

• model.safetensors: distilled HMamba weights
• config.json: architecture and preprocessing contract
• vocab.json: phone-to-ID mapping

Important limitations

• Input audio must have a reference English transcript.
• The model supports at most 52 aligned phone tokens.
• The original implementation requires CUDA-compatible mamba-ssm and Triton.
• Model outputs should not be interpreted as clinical or diagnostic results.

Feature contract

• GOP: 84 dimensions, normalized with mean 3.203 and standard deviation 4.045
• SSL order: Wav2Vec2 XLSR-53, HuBERT LL60K, WavLM Large
• SSL dimension: 1024 per model
• Raw input file: 17 dimensions; HMamba consumes duration plus the first seven energy statistics (indices 0 through 7)
• Phone scores use the 0-2 scale
• Word and utterance scores are multiplied by 5 for the 0-10 presentation scale