Nemotron 3 Diarization: who said what in your meeting?

Imagine a three-person meeting containing “I will send the draft tomorrow” followed by “I will check the numbers.” Even an accurate transcript is a poor task list without attribution. The next improvement is not necessarily a more fluent summary; it is preserving when each sentence occurred and which speaker said it.

ASR answers what was said; diarization answers which speaker was active when. Aligning the two creates a speaker-attributed transcript. If a local LLM³ later extracts decisions and tasks, retain timestamps so the reader can return to the original exchange.

Released on September 23, 2026, Nemotron 3 Diarization is an open-weight model with roughly 100 million parameters. It assigns generic speaker channels in arrival order and returns speaking intervals. A label such as speaker_0 is not a colleague's real name; assigning names requires a separate review or meeting metadata.

Two channels can be active at once when people overlap. This is not source separation: it does not extract a clean audio track for each voice. Detecting overlap does not automatically settle every word or its attribution. Conversations with more than eight speakers are outside the stated speaker limit.

A diarization model's name does not establish Korean transcription quality. An English meeting, a Korean discussion containing English product names, and a Japanese interview place different demands on ASR. Consider multilingual models such as Qwen3-ASR, Whisper or Nemotron 3.5, and check whether your chosen implementation supplies word timestamps.

Use a conversation resembling your work rather than only clean narration. Include interruptions, a distant speaker and someone returning after a long pause. Check three things separately: the words, stable speaker labels and overlapping intervals. Combining everything into one score makes it harder to see which stage needs attention⁴.

NVIDIA documents the NeMo Python path around Linux and NVIDIA GPUs. That does not make an NVIDIA card the only way to use the model. Its NeMo-Speech.cpp native runtime⁵ offers paths including Apple Silicon Metal, CPU⁶ and CUDA⁷. Check what your installed build actually supports with the doctor command.

For Mac and iPhone app development, Argmax Pro SDK 3 also supports Nemotron 3 Diarization. It is a separate product from the Python weights or C++ CLI⁸, with its own access and distribution terms. In every case, distinguish support from a measured time for your recording and device.

After installing NeMo-Speech.cpp for your operating system, open a fresh terminal. The first two commands inspect the environment and model catalog; pull fetches the weights in advance. Defaults can differ between releases, so the example explicitly selects the Nemotron 3 alias. If that name is missing from the catalog, check your installed version.

Use a short conversation recording you are authorized to process as meeting.wav. diarize returns speaker intervals; transcribe with --diar-model adds attribution to ASR output. Running the stages separately helps distinguish word errors from assignment errors. These commands have been checked against the documented CLI and model catalog, not timed on every device.

A streaming model waits for some incoming audio to interpret the surrounding context. The published 0.32, 0.64, 1.04 and 30.4 seconds describe that input buffer. They do not mean that an hour-long recording finishes in 30.4 seconds or that every device displays captions after 0.32 seconds. Compute, ASR and rendering add further time.

For a meeting that has already been recorded, immediate feedback may not matter. Start with a longer-context setting and reduce delay only if live display is actually needed. Do not copy configuration names blindly between NeMo Python and C++; use your runtime's presets and help output.

The launch article reports 14.72% DER in an initial VoiceArena evaluation of 139 English conversations totaling about 22 hours, scoring overlaps with no boundary collar. DER combines missed speech, false detections and speaker confusion; it is not a word-error score.

Calling this “85.28% transcription accuracy” would be wrong, as would treating it as the error rate for a Korean meeting.

The reported 15,113× throughput⁹ is striking, but it uses an RTX PRO 5000 with BF16¹⁰, batch 32 and torch.compile. It is not single-file latency on a personal computer or an ASR-inclusive result. We therefore have not applied it to the site's Mac mini or RTX 4090 simulations. A useful device comparison starts with matching audio, runtime and batch size¹¹.

Start with two to five minutes and manually note the speaker turns. Compare the transcript with the diarization around first appearances, interruptions and returning speakers. Replay passages that assign names or responsibilities, and do not turn an unresolved overlap into a definite statement by one person.

Use the same order when choosing hardware: check whether the models load, whether the job finishes faster than playback, and whether there is headroom with ASR running too. A task that only needs speaker labels does not require a large LLM to remain loaded. Keeping only the necessary stages may let your existing computer do the job.

Local execution does not remove the need for recording consent and a retention policy. The model is distributed under OpenMDW 1.1; apps and SDKs have separate terms. Initial downloads or updates may need the internet, so environments that must avoid external transfer should also check file locations and app synchronization settings.