Nvidia releases open model that labels up to eight speakers in live audio

Nvidia has released Nemotron 3 Diarization, a freely available model that identifies when each of up to eight people speaks in recorded or live audio. Paired with speech recognition software, it can produce transcripts with speaker labels, though those labels identify speakers only as anonymous entries such as "speaker_2." The model's audio buffer can be set between 30.4 and 0.32 seconds, with shorter buffers generally reducing accuracy. At a 1.04-second buffer, it cut the error rate of its predecessor by an average of 41 percent across eight test scenarios. More speakers, heavy background noise and reverb raise error rates.