README.md

---
license: cc-by-4.0
language:
- en
pretty_name: MondegreensEval
tags:
- speech
- automatic-speech-recognition
- audio
- bias
- hallucination
- whisper
task_categories:
- automatic-speech-recognition
size_categories:
- 1K<n<10K
dataset_info:
  features:
  - name: pair_id
    dtype: string
  - name: category
    dtype: string
  - name: source
    dtype: string
  - name: original_text
    dtype: string
  - name: mondegreen_text
    dtype: string
  - name: phoneme_original
    list: string
  - name: phoneme_mondegreen
    list: string
  - name: phoneme_edit_distance
    dtype: int64
  - name: phoneme_edit_distance_norm
    dtype: float64
  - name: condition
    dtype: string
  - name: audio_original
    dtype:
      audio:
        sampling_rate: 16000
  - name: audio_mondegreen
    dtype:
      audio:
        sampling_rate: 16000
  splits:
  - name: test
    num_bytes: 202606959
    num_examples: 1140
  download_size: 198883999
  dataset_size: 202606959
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# MondegreensEval

Audio companion dataset for **MondegreensEval: A Phonetic Benchmark for Measuring
Language-Model Bias in Automatic Speech Recognition** (ICML ML for Audio Workshop, 2026).

Code, evaluation pipeline, and per-model transcription/metric outputs:
https://github.com/soarhigh/mondegreenbench

Mondegreens — phonetically near-identical phrase pairs with distinct meanings — expose a
measurable failure mode in decoder-based ASR: the model's internal language-model prior
can override acoustic evidence and "correct" a spoken mondegreen back to the canonical
phrase (e.g. *"excuse me while I kiss this guy"* → *"excuse me while I kiss the sky"*).
This dataset provides the synthesized audio and phonetic annotations backing that
benchmark.

## Dataset composition

190 curated mondegreen pairs (114 song lyrics, 41 speech phrases, 35 liturgical/prayer
lines), each synthesized as both its canonical (*original*) and misheard (*mondegreen*)
text, at 6 acoustic conditions (clean + additive Gaussian white noise at SNR = 15, 10, 5,
0, −5 dB) — 190 × 6 = **1,140 rows** in the `test` split.

**Why both original and mondegreen audio are included:** the benchmark's headline metric
(MCR-mono) plays the *mondegreen* audio and checks whether the model wrongly outputs the
canonical phrase — but validating that this reflects a genuine language-model bias (not a
generic transcription artifact) requires the symmetric control: playing the *original*
audio and confirming the reverse confusion (MCR-orig) stays low. Both audio columns are
needed for either metric.

| Column | Type | Description |
|---|---|---|
| `pair_id` | string | Unique identifier, e.g. `sg_001` |
| `category` | string | One of `song_lyrics`, `speech_phrases`, `prayers` |
| `source` | string | Provenance (song/artist, phrase type, or liturgical source) |
| `original_text` | string | Canonical, higher-frequency phrase |
| `mondegreen_text` | string | Phonetically similar, lower-frequency misheard alternative |
| `phoneme_original` | list[string] | ARPAbet phonemes for `original_text` (stress-stripped) |
| `phoneme_mondegreen` | list[string] | ARPAbet phonemes for `mondegreen_text` (stress-stripped) |
| `phoneme_edit_distance` | int | Raw Levenshtein distance between the two phoneme sequences |
| `phoneme_edit_distance_norm` | float | Distance normalized by the longer sequence's length, in [0, 1] |
| `condition` | string | `clean`, `snr15`, `snr10`, `snr5`, `snr0`, or `snr-5` |
| `audio_original` | Audio (16 kHz mono) | TTS-synthesized audio of `original_text` under `condition` |
| `audio_mondegreen` | Audio (16 kHz mono) | TTS-synthesized audio of `mondegreen_text` under `condition` |

Phoneme distances span 0.038–1.000 (mean 0.332), grouped into four tiers used by the
paper's analysis: near-homophones (`< 0.10`, n=36 pairs), plausibly ambiguous
(`0.10–0.25`, n=64), weakly similar (`0.25–0.40`, n=16), and dissimilar (`> 0.40`, n=74).

## Synthesis details

Audio synthesized via Edge-TTS (`en-US-JennyNeural` voice), 16 kHz mono WAV. Noisy
variants generated by additive Gaussian white noise at the stated SNR, applied
deterministically (seed=42) to the clean signal. Phonemes derived from the CMU
Pronouncing Dictionary with a `g2p_en` fallback for out-of-vocabulary words.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("soarhigh/mondegreenbench", split="test")

# Only clean audio:
clean = ds.filter(lambda r: r["condition"] == "clean")

# Only the primary diagnostic tier (phonetically ambiguous, 0.10-0.25):
ambiguous = ds.filter(lambda r: 0.10 <= r["phoneme_edit_distance_norm"] < 0.25)
```

## Limitations

See the paper's Discussion/Limitations sections: all audio comes from a single TTS
engine and voice (no speaker variability or recorded human speech); noise is additive
Gaussian white noise, not an ecologically representative degradation (babble,
reverberation, codec artifacts); and each phrase has exactly one synthesized sample, not
resampled across multiple stochastic TTS seeds.

## Citation

<!-- TODO: fill in the final camera-ready author list / BibTeX once available. -->

```
@inproceedings{mondegreenseval2026,
  title     = {MondegreensEval: A Phonetic Benchmark for Measuring Language-Model Bias in Automatic Speech Recognition},
  author    = {Wan Ju Kang}
  booktitle = {ICML 2026 Workshop on Machine Learning for Audio},
  year      = {2026}
}
```

## License

CC BY 4.0. Note that the `song_lyrics` category quotes short excerpts of song lyrics for
research/benchmarking purposes; if you plan to redistribute or build commercial products
on this subset specifically, please review the underlying copyright status of the quoted
lyrics independently.