NVIDIA releases Nemotron 3.5 Transcribe for longer recordings and speaker attribution

NVIDIA released Nemotron 3.5 Transcribe for recorded audio with optional timestamps and speaker attribution.

NVIDIA published Nemotron 3.5 Transcribe on October 28 with downloadable checkpoints for speech recognition, translation and speaker attribution. The release targets recorded audio and supports input up to one hour, according to NVIDIA’s model collection. It is separate from the smaller Nemotron 3.5 ASR streaming model. That distinction matters to a team deciding between transcribing a completed recording and displaying live text while someone speaks. [NVIDIA model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B) · [NVIDIA speech collection](https://huggingface.co/collections/nvidia/nemotron-speech)

What the structured transcript contains

The model combines a speech recognition encoder with a speaker encoder, a timestamp alignment head and a language model. NVIDIA’s serving documentation describes an optional structured response from vLLM that adds start and end times for words, speaker tags and segments, a mapping between the transcript tags and the diarization output, and a field for alignment errors. An application needs to enable timestamp capture to request this information in the documented serving path. [NVIDIA serving documentation](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-FP8)

These fields support a reader that can jump from a transcript to the corresponding recording. They also make different failure modes visible. A transcript can have the right words and still assign a statement to the wrong participant. A segment can point to the right person but begin several seconds too early. A reviewer who needs to check disputed wording should have the underlying recording and be able to inspect both the speaker assignment and the timing.

Where to evaluate it first

NVIDIA lists base and lower precision checkpoints. Start with the version and runtime intended for production and record the hardware, model revision and transcription settings. A representative test set should include speech from the languages and accents the application expects, as well as interruptions, background noise and overlapping speakers. Compare word accuracy, speaker assignment and timestamp alignment separately. Translation deserves its own review because a translated transcript can read fluently while dropping the uncertainty or intent conveyed in the recording.

NVIDIA also released Granary v2 research materials alongside the model. The model and data releases can help teams inspect the task and reproduce selected evaluations. Teams still need separate permission and retention controls for recordings, especially when transcripts will be searchable or shared. [NVIDIA release overview](https://huggingface.co/blog/nvidia/nemotron-transcribe-granary)