README.md

---
language:
- en
license: apache-2.0
task_categories:
- automatic-speech-recognition
pretty_name: Nyra Disfluency Speech English
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# Nyra Disfluency Speech English

`nyrahealth/disfluency_speech_english` is an English speech dataset for evaluating **verbatim ASR**: models that should transcribe not only the intended words, but also fillers, cutoffs, repetitions, and sound events.

This dataset is based on the [AMAAI Lab DisfluencySpeech dataset](https://huggingface.co/datasets/amaai-lab/DisfluencySpeech) and reformatted for verbatim-transcription benchmarking with paired:

- `verbatim_transcript`: what the speaker actually said
- `intended_transcript`: a cleaned version of what the speaker meant to say

It is used by the [Nyra Verbatim Speech Benchmark](https://github.com/nyrahealth/nyra_verbatim_speech_benchmark), which evaluates verbatim ASR in detail and breaks errors down into fillers, sounds, cutoffs, repetitions, and intended-transcript failures.

For the exact convention definitions used by the benchmark, see:

- [Nyra Verbatim Speech Benchmark](https://github.com/nyrahealth/nyra_verbatim_speech_benchmark)
- [Verbatim transcript conventions](https://github.com/nyrahealth/nyra_verbatim_speech_benchmark?tab=readme-ov-file#verbatim-transcript-conventions)

## Source

This release is derived from:

- Dataset: [amaai-lab/DisfluencySpeech](https://huggingface.co/datasets/amaai-lab/DisfluencySpeech)
- Paper: [DisfluencySpeech -- Single-Speaker Conversational Speech Dataset with Paralanguage](https://arxiv.org/abs/2406.08820)

The original dataset provides annotated transcripts and several progressively cleaned transcript variants. This Nyra release converts the data into a format that is directly usable for verbatim-ASR evaluation with paired verbatim and intended references.

## Dataset Structure

The dataset contains `4,957` utterances and about `9.4` hours of audio.

Splits:

- `train`: `4,458`
- `validation`: `250`
- `test`: `249`

Features:

- `id`
- `audio`
- `duration_in_s`
- `split`
- `speaker`
- `verbatim_transcript`
- `intended_transcript`

## Transcription Conventions

### Verbatim Transcript

The `verbatim_transcript` follows a small set of explicit conventions so disfluencies and non-speech events can be evaluated consistently:

- **Cutoffs** use `*`, for example `th*` or `w*`
- **Fillers** are bracketed tags, primarily `[UH]` and `[UM]`
- **Sound events** are also bracketed tags, for example `[laughter]`, `[breath]`, or `[cough]`
- **Repetitions** are written as repeated words, not as separate tags
- Spoken words are otherwise written as they were said

Example:

```text
I mean we we [UH] should go on th* Thursday [laughter]
```

### Intended Transcript

The `intended_transcript` is the cleaned target for intended ASR. It removes disfluent material while preserving the speaker's meaning, including fillers, sound tags, repeated restarts, and cutoff fragments.

Example:

```text
verbatim: I mean we we [UH] should go on th* Thursday [laughter]
intended: we should go on Thursday
```

This makes the dataset suitable for evaluating both:

- verbatim transcription quality
- intended transcription quality

## Tag Analysis

The counts below were computed over the full dataset from `verbatim_transcript`.

Summary:

- utterances: `4,957`
- utterances with at least one bracketed tag or cutoff: `2,779`
- total bracketed tags: `4,039`
- total cutoff tokens: `582`

### Fillers

| Tag | Count |
| --- | ---: |
| `[UH]` | 2,568 |
| `[UM]` | 504 |

### Sound Tags

| Tag | Count |
| --- | ---: |
| `[laughter]` | 714 |
| `[breath]` | 105 |
| `[lipsmack]` | 59 |
| `[throatclearing]` | 55 |
| `[sigh]` | 18 |
| `[sniff]` | 12 |
| `[cough]` | 4 |

### Cutoffs

| Marker | Count |
| --- | ---: |
| `*` cutoff tokens | 582 |

These statistics are useful when interpreting benchmark results: fillers are common, laughter is the most frequent sound event, and cutoffs occur often enough to matter as a separate evaluation category.

## Benchmark Usage

This dataset is designed to be used with the [Nyra Verbatim Speech Benchmark](https://github.com/nyrahealth/nyra_verbatim_speech_benchmark).

That benchmark:

- derives gold disfluency labels automatically from the verbatim and intended transcript pair
- computes transcript metrics such as `vWER` and `iWER`
- computes event metrics for fillers, sounds, cutoffs, and repetitions
- provides detailed error analysis for verbatim ASR models

## Citation

If you use this dataset, please cite the original DisfluencySpeech paper:

```bibtex
@misc{wang2024disfluencyspeechsinglespeakerconversational,
  title={DisfluencySpeech -- Single-Speaker Conversational Speech Dataset with Paralanguage},
  author={Kyra Wang and Dorien Herremans},
  year={2024},
  eprint={2406.08820},
  archivePrefix={arXiv},
  primaryClass={eess.AS},
  url={https://arxiv.org/abs/2406.08820}
}
```