README.md

---
license: cc0-1.0
language:
- ja
task_categories:
- automatic-speech-recognition
- text-generation
tags:
- asr-error-correction
- japanese
- whisper
- phonetics
pretty_name: Mondegreen ASR error pairs
size_categories:
- 1K<n<10K
---

# Mondegreen ASR error pairs

`(ASR hypothesis, gold text)` pairs for Japanese ASR post-correction.

This build is **simulated** -- errors come from a phonetic corruption model, not from a real ASR system. It exists so the whole pipeline (gate training, benchmarks, figures, CI) is reproducible without a GPU. Treat every number derived from it as a stated assumption, not a measurement.

## How it was made

```
synthetic text
  -> phonetic corruption model (mondegreen.simulate)
  -> hypothesis
```

**No audio was involved in this build.** Errors were generated by perturbing the *reading* of each term with the confusion classes the distance function discounts (voicing, long vowels, geminates, moraic nasal) and re-rendering it the way ASR would -- as katakana, or as a homophone kanji spelling drawn from the bundled reading table. The acoustic condition fields below record which condition each record *would* correspond to, and are carried through for parity with the measured pipeline.

To rebuild this dataset with a real TTS -> Whisper round trip:

```bash
python scripts/harvest_errors.py --mode real --whisper-size small -n 2000
```

**No LLM judges correctness anywhere in this pipeline.**

- provenance: **simulated**
- pairs: 9000
- glossary terms used: 12000
- acoustic conditions: ['close/15.0', 'close/20.0', 'close/None', 'far/10.0', 'far/12.0', 'reverb/5.0']

## Source text and licence

| field | value |
| --- | --- |
| corpus | `synthetic` |
| licence | CC0-1.0 |
| verification | Generated by mondegreen.harvest.SentenceFactory; no third-party text. |
| url | — |

If you add a corpus, add it to `mondegreen.harvest.CORPUS_LICENSES` with a
verified licence first. The harvester refuses unknown corpora by design.

## Pathology labels

| label | 日本語 | how it is produced |
| --- | --- | --- |
| `term-phonetic` | 固有名詞の音韻的置換 | glossary term rendered as a homophone or near-homophone |
| `voicing` | 濁音・清音の取り違え | rendaku / devoicing slip inside a term |
| `long-vowel` | 長音の脱落・付加 | chouon added or dropped |
| `geminate` | 促音の脱落・付加 | sokuon added or dropped |
| `moraic-nasal` | 撥音の脱落 | moraic nasal swallowed, typically in far-field audio |
| `particle-drop` | 助詞の欠落 | unstressed particle lost |
| `word-drop` | 語の脱落 | short span deleted entirely |
| `number-unit` | 数字・単位の誤り | digit or counter substituted |
| `hallucination` | 定型の幻聴 | canned phrase emitted over silence or noise-only audio |

Observed counts in this build:

| label | count |
| --- | --- |
| `geminate` | 3217 |
| `hallucination` | 1080 |
| `long-vowel` | 5688 |
| `moraic-nasal` | 548 |
| `number-unit` | 821 |
| `particle-drop` | 450 |
| `term-phonetic` | 6093 |
| `voicing` | 3048 |
| `word-drop` | 259 |

## Fields

| field | meaning |
| --- | --- |
| `id` | stable record id |
| `gold` | the text that was spoken (exact) |
| `hypothesis` | what the ASR returned |
| `glossary_terms` | glossary surfaces occurring in `gold` |
| `error_types` | pathology labels |
| `speaker`, `speed`, `snr_db`, `room` | acoustic condition |
| `asr_model` | which ASR produced the hypothesis |
| `source_corpus`, `source_license` | provenance of the gold text |
| `split` | train / test (disjoint speakers, sentences and glossaries) |
| `provenance` | `measured` (real TTS+ASR) or `simulated` |

## Intended use

Training and evaluating **post-correction** systems. Not for training ASR models.

## Privacy

All names in this dataset are **synthetic**, generated by
`mondegreen.harvest.GlossaryBuilder`. No real person's voice or name was used, and
no real meeting audio exists anywhere in this pipeline.