Cutting-edge speaker detection, anywhere you go

Cutting-edge speaker detection, anywhere you go. Glimpse 1.2.7 brings NVIDIA's new Nemotron-3 Diarization model to the Library, running entirely on your Mac or PC. Every meeting, interview and lecture comes back with each line credited to the right person, with no upload, no account and no internet connection required.

More accurate than ever

On the AMI meeting corpus, 16 real meetings and more than 9 hours of audio, Nemotron-3 in Glimpse makes 42% fewer errors than the previous model.

• Four times fewer mix-ups. Lines credited to the wrong person drop from 3.6% to 0.9%.
• Up to 8 speakers. Twice as many voices, told apart in a single recording.
• Sharp, even in real time. With just one second of lookahead, it outperforms the previous model running with a full 30 seconds.

Remarkably fast

On Apple silicon, speaker detection runs up to 339 times faster than real time, more than three times faster than before. A 10 minute meeting is done in under two seconds. The model is smaller, too, at 106 MB.

Built for your device

Glimpse runs Nemotron-3 through transcribe.cpp, its on-device speech engine, using a GGUF version built for Glimpse. The port matches NVIDIA's reference implementation at every stage, and the compact 8-bit model scores the same as full precision. Your recordings never leave your computer.

The converted model is available to everyone on Hugging Face, with full benchmarks.

Availability

Speaker detection with Nemotron-3 is available now in Glimpse 1.2.7 for Mac and Windows. Existing users are upgraded automatically in the background. Speaker detection is part of the Library, included with every Glimpse license and the two-week free trial. Dictation remains free and unlimited.