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

A speech system can recognize words correctly yet transcribe the wrong person. Krisp's voice-isolation benchmark tests that problem by labeling stretches of each recording as the intended speaker talking, both speakers talking, or only a background speaker continuing. In the last case, the correct transcript is empty. Krisp reports that, across its 265 real recordings and 11 transcription configurations, pooled word error rate fell 73%, from 23.24% on raw audio to 6.24% with its default isolation model. Its phone-call control shows a cost: with almost no competing speech, error was 3.48% raw versus 3.91% with the default model, and a few individual recordings got markedly worse. These are vendor-reported results, and the benchmark separates background-speaker rejection from general transcription accuracy.