One Front-End, Two Tasks: How Speech Enhancement Affects ASR on Noisy Indian-Accented English and Speaker Diarization in Meetings

Speech enhancement is often placed before both automatic speech recognition (ASR) and diarization in meeting systems, on the assumption that enhanced audio helps both tasks. We tested this assumption by comparing raw audio with the output of five enhancers and with blends of the two, using two ASR models (Whisper, Parakeet) on Indian-accented English and two diarizers (pyannote, Sortformer) on AMI meetings. No enhancer helped: most made both tasks worse, and diarization even more so. For example, DeepFilterNet3 raised Whisper's word error rate at 0 dB from 14.6% to 24.5%, and pyannote's diarization error rate on clean audio from 17.1% to 28.8%. In pyannote, enhancement mainly damaged the segmentation stage. Blending at least 50% raw audio with the enhanced audio removed almost all of the damage. Enhancement makes audio sound better, but not better for these models.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24363179
Primary Topic
Speech Recognition and Synthesis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

One Front-End, Two Tasks: How Speech Enhancement Affects ASR on Noisy Indian-Accented English and Speaker Diarization in Meetings

Sathish Kumar Balasubramani
Zenodo (CERN European Organization for Nuclear Research)
Speech Recognition and Synthesis
preprint

One Front-End, Two Tasks: How Speech Enhancement Affects ASR on Noisy Indian-Accented English and Speaker Diarization in Meetings

Sathish Kumar Balasubramani
preprint en

Abstract

Speech enhancement is often placed before both automatic speech recognition (ASR) and diarization in meeting systems, on the assumption that enhanced audio helps both tasks. We tested this assumption by comparing raw audio with the output of five enhancers and with blends of the two, using two ASR models (Whisper, Parakeet) on Indian-accented English and two diarizers (pyannote, Sortformer) on AMI meetings. No enhancer helped: most made both tasks worse, and diarization even more so. For example, DeepFilterNet3 raised Whisper's word error rate at 0 dB from 14.6% to 24.5%, and pyannote's diarization error rate on clean audio from 17.1% to 28.8%. In pyannote, enhancement mainly damaged the segmentation stage. Blending at least 50% raw audio with the enhanced audio removed almost all of the damage. Enhancement makes audio sound better, but not better for these models.

Zenodo (CERN European Organization for Nuclear Research)
Speech Recognition and Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon