I am thinking about replacing pyannote with the new Nemotron 3 Diarization from NVIDIA. It claims to have a much lower DER (Diarization Error Rate), detecting up to 8 speaker (pyannote: ~4) in one audio file, and handle overlapping speech especially well. This is all to be tested and verified.
There is a GGML version out that is supported by https://github.com/0xShug0/audio.cpp. It looks quite experimental at the moment. We'll need to wait until its matured. But audio.cpp should provide CUDA, HIP/ROCm, Vulkan, Metal, and CPU backends, which would be ideal for noScribe. The GGML Modell is 199MB in size - much bigger than pyannote, but still reasonable compared to the whisper models we ship.
I am thinking about replacing pyannote with the new Nemotron 3 Diarization from NVIDIA. It claims to have a much lower DER (Diarization Error Rate), detecting up to 8 speaker (pyannote: ~4) in one audio file, and handle overlapping speech especially well. This is all to be tested and verified.
There is a GGML version out that is supported by https://github.com/0xShug0/audio.cpp. It looks quite experimental at the moment. We'll need to wait until its matured. But audio.cpp should provide CUDA, HIP/ROCm, Vulkan, Metal, and CPU backends, which would be ideal for noScribe. The GGML Modell is 199MB in size - much bigger than pyannote, but still reasonable compared to the whisper models we ship.