README.md

---
pretty_name: "BLEEP: Broadcast Language Elicitation and Evaluation for Profanity"
language:
  - en
language_details: "en-US, en-GB"
multilinguality:
  - monolingual
license: other
license_name: bleep-dua
license_link: LICENSE
task_categories:
  - audio-classification
task_ids:
  - keyword-spotting
annotations_creators:
  - other
language_creators:
  - crowdsourced
source_datasets:
  - original
size_categories:
  - 1K<n<10K
tags:
  - audio
  - keyword-spotting
  - profanity
  - hard-negatives
  - accented-english
extra_gated_heading: "Request access to BLEEP"
extra_gated_description: >-
  Access is granted to individuals, not organisations, and is reviewed manually.
  Please allow a few days for a decision.
extra_gated_button_content: "Accept the DUA and request access"
extra_gated_prompt: >-
  BLEEP is released for non-commercial academic research only, under a Data Use
  Agreement. You agree not to (a) attempt to re-identify any speaker; (b) use the
  recordings for voice cloning, speaker verification, surveillance, or harassment;
  (c) redistribute the audio or metadata; (d) use the data commercially. Commercial
  licences are separate and administered by the maintainer. The dataset contains
  strong profanity and sexually explicit lexical items.
extra_gated_fields:
  Full name: text
  Institutional email: text
  Affiliation: text
  Country: country
  Intended research use: text
  I have read and agree to the BLEEP Data Use Agreement: checkbox
  I will not attempt to re-identify speakers or use the data for voice cloning: checkbox
  I will not redistribute the data or use it commercially: checkbox
---

# BLEEP — Broadcast Language Elicitation and Evaluation for Profanity

**v2.0** · 8,312 clips · 86 speakers · English (US/UK) · 16 kHz mono · 4.62 h

Isolated-word English corpus for profanity speech research. 20 profanity keywords and 29 hard
negatives — minimal-pair confusables selected by CMUdict phoneme edit distance and SUBTLEX-US/UK
frequency. Each speaker recorded all 49 words twice in one session, once in a **neutral** and
once in an **expressive** register.

> ⚠️ **Content warning** — every clip is a spoken profanity or a near-homophone of one.
> 🔒 **Gated, research-only, non-commercial** under a Data Use Agreement. 
## Composition

| | |
|---|---|
| Clips / speakers | 8,312 / 86 (`ID00001`–`ID00086`), mean 96.7 each |
| Vocabulary | 49 words — 20 profanity, 29 hard negatives |
| Registers | `neutral` 4,169 · `expressive` 4,143 |
| Labels | 3,384 `profanity` / 4,928 `non_profanity` |
| Audio | 16 kHz mono 16-bit WAV, fixed **2.000 s**, word onset at 200 ms |
| Countries | US 49 / UK 37 · collected 20–30 June 2026 via Prolific |

### Word list

12 base words, 8 morphological variants, 29 hard negatives — 49 in total. Hard negatives are
at phoneme edit distance *d* = 1 from their base word in every case except `mustard` (*d* = 2).

| Base word | Variants | Hard negatives |
|---|---|---|
| fuck | fucking, fucked, fucker | duck, luck, suck, buck |
| shit | bullshit, shitty | sit, shot, shut, ship, sheet |
| cunt | — | count, hunt |
| piss | — | miss, kiss, piece, pick |
| tits | — | bits, hits, sits, tips |
| motherfucker | motherfucking | *none* |
| cocksucker | — | *none* |
| bitch | bitches, bitching | beach, pitch, witch, ditch |
| ass | — | gas, pass, mass, lass |
| pussy | — | pushy |
| bastard | — | mustard |
| asshole | — | *none* |

Base words and variants are `label` = `profanity`; hard negatives are `label` = `non_profanity`.
All 49 words are recorded in both registers. `cocksucker`, `motherfucker` and `asshole` have no
hard negatives — no CMUdict+SUBTLEX neighbours exist at *d* ≤ 2.

**Not included:** silence/background and unknown/filler classes, connected speech, codec- or
noise-processed variants, TTS augmentation, phone alignments.

## Files

```
clips/ID000NN/ID000NN_{word}_{register}.wav
metadata.csv       # HF loader index (file_name + manifest columns)
manifest.csv       # clip_path, speaker_id, word, register, tier, label,
                   # word_onset_s, word_offset_s, clip_ms
demographics.csv   # speaker_id, gender, age_range, background, accent_dialect,
                   # country, languages, english_acquisition_age, media_exposure
```

`label` ∈ {`profanity`, `non_profanity`}. `word_onset_s` and `word_offset_s` are in the source
recording's timebase. Clips are cut at `onset − 200 ms` with fixed 2 s length; within a clip the
word spans `[200 ms, 200 + (offset − onset)]`. Windows extending past the source are zero-padded.

Demographics are self-reported and categorical. Free-text region was dropped and generalised to
country. `accent_dialect` is pipe-delimited multi-select; an empty value is a non-response.

## Speakers

| Field | Distribution (n = 86) |
|---|---|
| `gender` | Female 45 · Male 39 · Non-binary 2 |
| `age_range` | 36–45: 26 · 26–35: 23 · 46–55: 16 · 18–25: 12 · 56–65: 8 · 65+: 1 |
| `country` | US 49 · UK 37 |
| `background` | white 48 · black 20 · hispanic/latino 5 · south asian 5 · MENA 2 · mixed 2 · asian 1 · unstated 3 |
| `languages` | english_only 74 · english_plus 8 · unstated 4 |
| `english_acquisition_age` | from birth 72 · before 5: 6 · age 5–10: 2 · age 18+: 1 · unstated 5 |
| `media_exposure` | mostly US 43 · mostly UK 19 · mixed 17 · unstated 7 |

`accent_dialect` is multi-select: 116 tags across 86 speakers in 29 distinct combinations. The
column below sums to 116, not 86.

| Accent / dialect | Speakers |
|---|---:|
| General American | 26 |
| Southern US | 16 |
| Standard Southern British / RP-like | 12 |
| African American English | 10 |
| New York / Northeast US | 9 |
| Northern England | 9 |
| Western US / California | 6 |
| Midlands English | 6 |
| London Estuary | 5 |
| Midwest US | 4 |
| Scottish English | 4 |
| Hispanic/Latino English | 3 |
| Welsh English | 2 |
| Northern Irish English | 2 |
| Asian American English | 1 |
| Unknown | 1 |

## Collection and QC

Recorded through a web app (Chrome desktop) in ~8–10 min: mic check, 49 words neutral, 49 words
expressive, demographic questionnaire. Prompt order was randomised per speaker. Each trial
captured 2.5 s at 48 kHz — a 500 ms lead-in followed by a 2 s speak window. Participants could
skip any word or stop early without affecting payment. The expressive prompt read "as in a TV or
film scene, never shouted or directed at anyone".

QC was run independently of alignment: Silero VAD localisation (threshold 0.5), duration,
clipping and segmental-SNR checks (flag only), and Whisper large-v3 with a primed initial prompt
for substitution detection. A recording was rejected only where VAD found no word and Whisper
did not confirm it. Speakers retaining fewer than 60 clips were excluded.

Of 9,286 recordings: **8,312 accepted (89.5%)**, 865 dropped with an excluded speaker
(9 speakers), 109 rejected individually.

Alignment used MFA 3.x with the `english_us_arpa` acoustic model and a supplementary profanity
dictionary.

## Uses

**Permitted:** profanity keyword-spotting research; phonetic confusability and false-alarm
benchmarking; research on ASR profanity suppression.

**Prohibited under the DUA:** voice cloning, TTS, and speaker verification on any released
speaker; speaker re-identification; surveillance and harassment; commercial use; redistribution;
training offensive-content generators.

## Consent — what participants were shown

Participants read an information sheet and ticked each item separately before recording.
Responses were logged with the document version and a timestamp.

Required consent items:

| # | Item |
|---|---|
| 1 | Aged 18+, has read and understood the document |
| 2 | Understands the recordings involve profane language and is willing to produce it; may skip, stop, or withdraw |
| 3 | Explicit consent to biometric processing under UK GDPR Art. 9(2)(a), including transfer to the controller in India and to DUA recipients |
| 4 | Grants a perpetual, worldwide, **non-exclusive** licence |
| 5 | Understands the dataset may be commercially licensed (excluding synthetic derivatives) with no further compensation |
| 6 | Understands and accepts the re-identification risk |
| 7 | Consents to demographic information being collected, stored alongside the recordings, and disclosed to approved researchers and reviewers |

## Licence

Research-only, non-commercial, gated release, per person, under a Data Use Agreement — not
CC-BY, and redistribution is prohibited. The maintainer holds a non-exclusive licence from
each speaker and separately administers commercial licences covering the original recordings
only, excluding synthetic derivatives. Re-releasing openly under CC-BY would require
re-consent, and an erasure request can remove a speaker from a released version.

**Maintainer:** Ritin Raveendran Kasthuri — `contact@methodosprojects.org`

## Usage

```python
from datasets import load_dataset, Audio
import numpy as np

# Gated: requires an approved access request and `hf auth login`.
# The corpus is one undivided partition, named "train" by convention. No splits are shipped.
ds = load_dataset("<org>/BLEEP", split="train").cast_column("audio", Audio(sampling_rate=16_000))

# Speaker-disjoint split.
spk = sorted(set(ds["speaker_id"]))
np.random.default_rng(0).shuffle(spk)
n = len(spk)
test, dev = set(spk[: n // 7]), set(spk[n // 7 : 2 * n // 7])
train = set(spk) - test - dev
splits = {k: ds.filter(lambda r, s=s: r["speaker_id"] in s)
          for k, s in [("train", train), ("dev", dev), ("test", test)]}

# Word extent within a clip.
word_span_ms = lambda r: (200.0, 200.0 + (r["word_offset_s"] - r["word_onset_s"]) * 1000)
```

## Citation

```bibtex
@misc{kasthuri2026bleep,
  title   = {{BLEEP}: Broadcast Language Elicitation and Evaluation for Profanity},
  author  = {Kasthuri, Ritin Raveendran},
  year    = {2026},
  version = {2.0},
  note    = {English profanity keyword-spotting corpus with systematic phonetic hard
             negatives. Gated research-only release under a Data Use Agreement.},
  howpublished = {\url{https://huggingface.co/datasets/<org>/BLEEP}}
}
```