README.md

metadata

license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
  - question-answering
  - feature-extraction
language:
  - en
tags:
  - speechgr
  - slue-sqa5
  - hubert
  - discrete-units
  - kmeans

SLUE-SQA-5 HuBERT Layer-22 K=500 Discrete Units

Packed discrete-unit files for SpeechGR experiments on SLUE-SQA-5.

The units were produced with HuBERT layer 22 and a K=500 k-means model, then deduplicated with consecutive counts retained. The packed format avoids one .code and .cnt file per utterance.

Files

• documents.npz: packed document/passage units
• train.npz: packed train question units
• validation.npz: packed validation question units
• test.npz: packed test question units
• verified_test.npz: packed verified test question units

Each archive contains:

• ids: record ids
• codes: all unit sequences concatenated
• code_offsets, code_lengths: offsets into codes
• counts: all consecutive-run counts concatenated
• count_offsets, count_lengths: offsets into counts
• text: optional transcript/passage metadata when available
• doc_ids: question-to-document ids for question splits when available

Loading

Use unit_store.PackedUnitStore from the SpeechGR repository:

from unit_store import PackedUnitStore

store = PackedUnitStore("train.npz")
record_id = store.ids[0]
codes = store.get_code(record_id)
counts = store.get_counts(record_id)