README.md

---
license: apache-2.0
tags: [automatic-speech-recognition, librispeech, error-analysis]
---

# Worst-100 test-clean clips — audio, transcripts, and the vocabulary finding

The 100 LibriSpeech test-clean clips where the block-4 production model (4.60% WER) made
the most word errors — with audio embedded so the failures can be *listened to*, plus the
model's transcript next to the reference for each clip.

## The finding this dataset produced

**47% of the word errors in these clips are on words that never appeared in the 30-hour
training vocabulary at all** (20,066 distinct words seen). The model was not mishearing
these words — it had never met them. Repeat offenders are book-specific names (rodolfo,
montrose, boolooroo). Full breakdown in `worst100_oov_analysis.json`.

This measurement is the primary justification for the 100-hour training phase: a coverage
problem calls for coverage. See `Diffusion-ASR/pipeline-code-and-docs` for the full recipe
and evidence record.


## How to read this dataset (verified 2026-08-11)

- **Ranking basis:** clips are ranked by **absolute word-error count** on
  **Whisper-normalized** text — not by WER percentage. A long clip with 20 errors at 23%
  WER outranks a short clip at 60%.
- **Tie band:** the rank-100 cutoff is 4 errors, and 66 clips tie at exactly 4 — which of
  them appear here is arbitrary among equals.
- **`substitution_pairs` are raw alignment output**, not claimed confusions: across a
  garbled stretch, positional pairing produces items like `have→all` that carry no
  meaning individually. Use `error_classes` for interpretation.
- Some references contain **genuine dialect spellings** (e.g. `allers`) — that is the
  LibriSpeech ground truth, not a transcription error.
- Integrity verification: all 100 references byte-match the official test-clean
  transcripts; all 100 hypotheses byte-match the model's recorded decode outputs; stored
  error counts reproduce exactly under the stated normalization.