Nvidia drops a free 100M-parameter model that identifies up to eight speakers in real time
decoder ·
Nvidia has released Nemotron 3 Diarization, a free AI model with approximately 100 million parameters designed to identify speakers in real-time audio. The system can distinguish up to eight different voices and detect instances of overlapping speech within conversations. It is compatible with both recorded files and live audio streams, often pairing with speech recognition tools like Parakeet to generate labeled transcripts. In the VoiceArena Diarization Benchmark v1, the model achieved a Diarization Error Rate (DER) of 14.7%, placing it at the top of current rankings. This performance represents a significant improvement over its predecessor, Streaming Sortformer, reducing error rates by an average of 41 percent across various test scenarios. While the model handles standard conditions well, accuracy may decrease when dealing with heavy background noise, reverb, or more than eight participants. Users can adjust the audio buffer settings between 0.32 and 30.4 seconds, though shorter buffers generally result in lower precision.
decoder