Ontario-based VIQ Solutions and Canadian venture studio AXL say they have achieved 97 to 99 percent accuracy in identifying speakers in multi-person transcripts, addressing a persistent problem known as diarization.
VIQ says existing speech engines still struggle with short utterances, similar voices, and overlapping speech, which can lead to words being assigned to the wrong person. Because multi-speaker recordings account for roughly 95 percent of VIQ’s speech output, the companies made diarization a priority. The new capability will enter pilot testing before a broader rollout, helping users review drafts faster with fewer speaker corrections and names automatically assigned more reliably.
Want to know more? Check out the source code here.