Get in touch

Research

Speech and audio research that grew out of product work, plus side projects on how learned representations line up with human perception.

Publications

Accepted · IEEE SLT 2026 Demo TrackTo be presented · Palermo, Dec 2026

BLEEP: A Broadcast Audio Dataset for Word-Level Profanity Detection

R. R. Kasthuri

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.

8,312word-level clips
86US and UK speakers
29hard negatives
0.11–0.13ΔCER, SSL models
Dataset & codePaper page →

Talks and notes

  • BLEEP at IEEE SLT (accepted, to be presented)
    TalkPalermo, Italy

    A word-level English speech dataset of 8,312 clips from 86 US and UK speakers, built to measure how often speech recognisers miss profanity in broadcast audio. Accepted at the IEEE SLT 2026 Demo Track, to be presented in Palermo in December.

  • EEG × ResNet-50: a time-resolved RSA pilot
    NotesTHINGS-EEG, strongest alignment near 120 ms

    Compares human EEG responses with pixel, untrained and pretrained ResNet-50 representations using time-resolved RSA. Pretrained features align best, peaking around 120 ms after the image appears, with mid-level layers strongest.

  • TalkTechnical talk, LG

    Internal technical talk at LG Soft India covering the fundamentals of audio, how audio is represented for machine learning, and the CNN- and Transformer-based architectures behind modern audio ML.

  • TalkTechnical talk, LG

    Internal technical talk at LG Soft India on using a model's embedding space to understand what supervised models learn and how they separate the classes they're trained on.