Show HN: Open benchmark for STT when a second person is talking (265 recordings)

Category: ai-ml

Tags: speech-recognition, benchmark, voice-isolation, dataset, stt

Score: 7.5/10 (Innovation: 7, Technical: 8, Documentation: 7, Utility: 8)

This project provides an open benchmark dataset and evaluation harness for measuring speech-to-text (STT) performance when a competing background voice is present. It includes 265 real-world recordings across work and call center scenarios, with hand-labeled segments and results from eleven STT systems, addressing a critical gap in evaluating voice isolation and diarization. The benchmark enables reproducible comparison and drives improvement in handling overlapping speech, a common failure mode for voice agents.

Target audience: Speech recognition researchers, ML engineers, voice AI developers, and product teams building voice interfaces

Repository: https://krisp.ai/blog/voice-isolation-benchmark/ · Python · Apache-2.0 · 513 stars

View on Hacker News