An evaluation-only preview licence gates NVIDIA's top-ranked diarization model

Call the model and you get a floating point tensor of shape [T, 8]: T time steps by eight speaker channels. Each value is the probability that a speaker is active at that moment. The default stride is 10 ms, and it can be set to any other multiple of 10 ms ^[18]. Words come from elsewhere. A product needs a separate speech-recognition system, then has to align its text against these timestamps ^[8]. The Hugging Face post is direct about what the channels mean: "These are anonymous labels, not real-world identities: the model can report that speaker_2 spoke from one timestamp to another, but it does not determine that speaker_2 is a particular person." ^[20] The front of the pipeline is easy to reason about. Audio arrives at 16 kHz, single channel. It becomes Mel-spectrogram features on a 10 ms frame step, gets stacked eight-to-one into 80 ms frames, and feeds a 31-layer Transformer encoder with rotary positional embeddings. A Conv1D layer above the encoder upsamples predictions back to input-feature resolution ^[17][18]. The harder problem is keeping a label attached to a voice. A streaming model sees one short chunk and limited context, and without a memory mechanism the speaker on output channel one can land on channel two in the next chunk ^[26]. Nemotron orders output speakers by first appearance: first new voice, first channel, and the per-chunk permutation problem goes away ^[19]. NVIDIA senior engineer Francesco Ciannella wrote the release article with Maryam Motamedi, Taejin Park, Ivan Medennikov and Adi Margolin. It states the target as keeping labels stable as a conversation arrives in chunks, including when participants interrupt one another ^[25][27]. The 4.58-point gap to the next-ranked system is a 23.7% relative reduction ^[1], and the conversations behind it average roughly nine and a half minutes each ^[2]. DER sums missed speech, false detections and speaker confusion ^[9]. A strong total can sit on top of the error a particular product cares about, such as an overlapping utterance assigned to the wrong participant ^[10]. The other headline figure comes from a different test. Against the four-speaker Streaming Sortformer baseline, the diar_streaming_sortformer_4spk-v2.1 checkpoint, NVIDIA reports a 41.0% average relative DER reduction across eight public evaluation conditions at the 1.04-second input buffer ^[11][13]. That figure is the unweighted mean of the eight relative improvements ^[12]. For it to transfer, your audio has to resemble those conditions and you have to run at 1.04 seconds. An unweighted mean lets a large gain on a condition you never encounter pay for a small gain on the one you do. The buffer setting is the next choice. NVIDIA recommends 30.4, 1.04, 0.64 and 0.32 seconds ^[14], a 95x spread between the longest and shortest ^[4], and says lower buffering generally reduces accuracy and throughput ^[16]. The 0.32-second setting is the floor of those recommendations, and computation, network transport, speech recognition and application processing all add time on top ^[15]. On the data side, adding David AI's licensed audio moved compound DER from 11.19% to 10.42% at both the offline-style and ultra-low-latency operating points ^[22], a 6.9% relative reduction ^[3]. The licensed material included multispeaker-annotated real conversations plus source audio for simulated English and multilingual mixtures spanning 21 languages ^[21]. NVIDIA did not describe the commercial terms or the full scope of the agreement ^[23]. David AI was founded in 2024 by Tomer Cohen and Ben Wiley, according to its Y Combinator profile ^[24]. The licence split comes from runtimewire.com's reading of the preview page against NVIDIA's own announcement, which uses the word open-weight ^[6]. Evaluation-only permits internal testing, which is the work a team should do first anyway: the leaderboard was scored on 139 English conversations, and your own audio is the only test that answers your question ^[4].