BLEEP: A Broadcast Audio Dataset for Word-Level Profanity Detection
Ritin Raveendran Kasthuri · Independent research
Abstract
Speech recognisers trained on clean read speech under-recognise profanity, which makes live broadcast filtering unreliable. BLEEP is a word-level English dataset with phonetically confusable "hard negatives", recorded in neutral and expressive registers, with a benchmark across SSL and Whisper models.
Illustration: one profane word located and muted inside a speech clip.
- word-level clips
- 8,312
- speakers kept (of 95)
- 86
- terms + spoken variants
- 12 + 8
- hard negatives
- 29
- ASR models benchmarked
- 5
- ΔCER for SSL models
- 0.11–0.13
Why it exists
Live broadcast filtering depends on speech recognisers catching profanity at the word level. Models trained on clean read speech systematically under-recognise exactly those words, and there was no word-level dataset built for broadcast conditions to measure it.
BLEEP is that dataset: 8,312 word-level English clips from 86 US and UK speakers, with a benchmark across self-supervised and Whisper models and a metric that isolates how much worse models do on profanity than on matched clean words.
It started as a finding from live profanity filtering on LG webOS TVs.
How it was built
Vocabulary built around broadcast rules
12 base profanity terms and 8 spoken variants, tiered by FCC and Ofcom severity, plus 29 phonetically confusable clean "hard negatives" chosen by phoneme distance in CMUdict and filtered by SUBTLEX-US/UK frequency so they are words common in broadcast speech.
Crowdsourced recording
A web-based elicitation interface used with 95 speakers recruited on Prolific: consent, microphone check, demographic metadata (accent, region, age, gender), and randomised neutral and expressive recording blocks.
Automated quality control
Silero VAD for speech localisation, SNR estimated against per-speaker silence recordings, and Whisper large-v3 content verification, followed by manual review. 9 speakers and 109 clips were excluded.
Fixed-onset alignment
A version aligned with the Montreal Forced Aligner and a custom profanity pronunciation dictionary, removing variation from reaction time and recording latency.
Benchmark and ΔCER
wav2vec2-base/large, HuBERT-large and Whisper-tiny/base benchmarked with a ΔCER profanity-penalty metric. Clean-speech SSL models err 0.11–0.13 CER more on profanity than on matched clean words. Whisper's gap is near zero.
Ethics and release
Consent obtained for voice as biometric data, with disclosure of re-identification risk. Dataset and code are released under a gated, research-only data use agreement.
Dataset and code
The dataset and code are released under a gated, research-only data use agreement. The paper will be linked here once the proceedings are published.