NVIDIA Releases Nemotron 3.5 Transcribe Speech LLM

This model supports 40 languages and can process recordings up to one hour in length. Users can select output styles such as verbatim, written form, speaker-attributed, contextual, and translated output through text instructions.

Nemotron 3.5 Transcribe is positioned as a transcription specialist, not a general-purpose assistant. It integrates an audio encoder with an LLM decoder to convert audio and instructions into text. The model's architecture includes a Parallel Expert Encoder with a Transformer ASR encoder and a frozen Nemotron 3 Diarization model, which tracks up to 8 speakers.

The LLM backbone is Nemotron 3.5 Lightning 30B-A3B, a hybrid Mamba-Transformer Mixture of Experts (MoE) model. The total size is 34 billion parameters with approximately 4.3 billion active parameters. It features a 65,536-token context window and runs on a single H100 80 GB GPU in BF16 precision. Checkpoints are available in FP8 and NVFP4 formats, with the latter running on an RTX 5090.

In performance, Nemotron 3.5 Transcribe is noted as the best open-weights model on English and multilingual Open ASR Leaderboards, achieving 4.71% mean Word Error Rate (WER) on English datasets. However, top commercial APIs still maintain a lead. The model's accuracy on Mandarin far-field meetings was 34.95% cpCER.

A key benefit is its ability to provide transcripts, speaker labels, timestamps, and translation from a single open model, though it is offline-only and has weak Mandarin meeting accuracy.