NVIDIA Nemotron 3.5 Transcribe: A Complete Guide to the Speech AI Model

NVIDIA Nemotron 3.5 Transcribe 30B A4B is an open-weight, offline speech-and-audio model capable of transcribing recordings, translating speech, identifying speakers, and answering questions about audio across 40 languages. It was developed by NVIDIA and serves various use cases, such as verbatim transcription, removing filler words, and creating speaker-attributed transcripts.

The model's architecture combines a Mamba-2–Transformer hybrid mixture-of-experts language model with a Parallel Expert Encoder, including a Transformer ASR encoder and speaker encoder. The checkpoint contains 34B total parameters, with approximately 4.3B active. It supports context lengths of up to 65,536 tokens and audio recordings up to one hour in length. NVIDIA provides the model checkpoint in BF16 format, with FP8 and NVFP4 variants also available.

To deploy NVIDIA Nemotron 3.5 Transcribe, a user must install vLLM 0.32.0 and the latest NeMo Speech main branch. Then, they can use the provided Python code snippet to send a WAV audio file as a base64 data URL and receive the transcription output. The model supports various configurations, such as requesting verbatim or cleaned text, preserving or omitting fillers, and including speaker tags.

The model does not support real-time streaming and is intended for offline processing. It was evaluated using benchmark datasets for English, long-form, and multilingual ASR, with mean WER scores of 4.71, 10.98, and 4.62, respectively. While these scores provide a comparison to other open-weight models, they do not guarantee performance on all workloads.

NVIDIA Nemotron 3.5 Transcribe is released under the OpenMDW-1.1 license and is supported by NVIDIA's NIM serving measurements on a single RTX PRO 6000 Blackwell Server Edition GPU. However, these optimizations are not included in the open-source release and should not be considered standard performance metrics.