NVIDIA's Nemotron 3.5 Transcribe, released October 28, transcribes recorded audio up to one hour in 40 languages. It accepts instructions and can label speakers and return word timestamps.
The big change
- A transcript can serve different purposes: a verbatim record preserves hesitations, while a readable version can omit them. A team needs to choose the form its readers require.
- Speaker labels help people follow a conversation, but do not establish a person's real-world identity. Check attributed speech against the recording and names against reliable records.
- A benchmark compares output with a defined reference. Its score is useful evidence for that test, while the accuracy of a particular meeting depends on the audio and task.
From recording to transcript
NVIDIA combines a speech encoder and frozen Nemotron 3 Diarization component with the Nemotron 3.5 Lightning language model. Its model card describes prompts for style, vocabulary and speaker tags; timestamps are aligned after generation. Context terms are hints, not evidence of identity.
The model uses the OpenMDW 1.1 license. NVIDIA documents an H100 80 GB BF16 profile, L40S 48 GB FP8 and RTX 5090 NVFP4. BIG CHANGE reviewed the documentation and did not run the model.
Reading the results
NVIDIA reports 4.71% mean word error on nine English leaderboard sets; the next listed open-weight model scored 4.94%. These vendor-reported results are test-specific.
The Open ASR Leaderboard methodology measures substitutions, insertions and deletions against a reference and normalizes text before scoring. That score does not by itself establish whether speaker labels or a requested transcript style are correct.
BIG CHANGE previously covered the standalone diarization model. This is a separate release.
