noisekit generates degraded speech corpora for ASR testing
noisekit is a CLI for turning clean, labeled speech datasets into noisy, benchmark-ready corpora for speech-to-text evaluation. It applies realistic degradations like telecom codec artifacts, ambient noise, reverberation, clipping, and low-bitrate compression, then outputs an AudioFolder-compatible dataset with original transcripts plus signal-quality metadata such as PESQ, SNR, and NISQA. The pitch is practical: if you can’t label your real production audio, simulate the failure modes you actually care about and compare STT vendors on something closer to your deployment conditions.
Strong utility for teams doing serious ASR vendor selection, but the value depends on how closely the simulated degradations match your real audio pipeline.
- –The core problem is real: clean public corpora are a poor proxy for phone calls, contact-center audio, and other messy production inputs.
- –The output format is useful: HuggingFace AudioFolder compatibility means this slots cleanly into existing evaluation workflows.
- –The preset design is the main differentiator: telecom, noise, reverb, clipping, and compound chains map to common real-world failure modes better than generic “add white noise” demos.
- –The biggest caveat is also stated by the author: simulated degradation is not the same as genuine production audio, so it is best used for relative benchmarking and stress testing, not as a final truth source.
DISCOVERED
45d ago
2026-05-28
PUBLISHED
46d ago
2026-05-27
RELEVANCE
AUTHOR
Karamouche