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

Signup Write New Story NVIDIA Nemotron 3.5 Transcribe: A Complete Guide to the Speech AI Model by aimodels44 by aimodels44 | @ aimodels44 Among other things, launching AIModels.fyi... Find the right AI model for your project - https://aimodels.fyi

Among other things, launching AIModels.fyi... Find the right AI model for your project - https://aimodels.fyi

NVIDIA Nemotron 3.5 Transcribe 30B A4B is an open-weight, offline speech-and-audio model that turns recordings and text instructions into text. It supports transcription, translation, speaker-attributed and contextual transcription, timestamps, language identification, and audio question answering across 40 languages. The model is maintained by nvidia.

This is an offline model, not a real-time streaming ASR service; internal audio chunking does not change that. The stated one-H100 80 GB deployment requirement is substantial. Published benchmark figures are reported for selected datasets and tasks, so they do not establish performance on every language, domain, or recording condition.

The supplied results report mean WER of 4.71 for English ASR, 10.98 for long-form ASR, and 4.62 for multilingual ASR across the listed benchmark sets; they identify Qwen3-ASR-1.7B and Parakeet-TDT-0.6B-v3 as next-best open-weight models on those respective evaluations. These are benchmark-specific reported comparisons, not a guarantee for other workloads. The related Nemotron Speech Streaming English 0.6B model is described as supporting streaming, unlike this offline model. The supplied Nemotron 3.5 ASR Streaming 0.6B listing says it is in internal evaluation and does not provide enough public detail here for a capability comparison. FP8 and NVFP4 are precision variants of this Transcribe checkpoint, not separate task descriptions.

The checkpoint has 34B total parameters, with approximately 4.3B active including the audio encoder. Its architecture is a Mamba-2–Transformer hybrid mixture-of-experts language model with a Parallel Expert Encoder comprising a Transformer ASR encoder and speaker encoder. This exact checkpoint is the BF16 reference; FP8 and NVFP4 variants are also listed.

Context length is 65,536 tokens, and supported recordings can be up to one hour. The README gives one H100 80 GB as a deployment configuration. Native Multi-Token Prediction is supported, with an optional DFlash2 draft model.

Reasoning can be enabled or disabled through the chat template. Recommended settings are temperature 0 and reasoning off for transcription, translation, and language identification; for audio understanding, temperature 0.6, top_p 0.95, and reasoning on. The stated license is OpenMDW-1.1.

The README also reports optimized NIM serving measurements on one RTX PRO 6000 Blackwell Server Edition GPU, but those optimizations are not part of the open-source release and should not be treated as performance measurements for this checkpoint's standard open-source serving setup.

Accepts audio recordings together with text instructions and optional context. The README's serving example sends WAV audio as a base64 data URL; its Transformers example uses an audio sample URL. Instructions can specify language, verbatim or cleaned wording, formatting, context terms, translation target, speaker tags, or an audio question. Word timestamps can be requested; the README notes that vLLM word timestamps require the V1 model runner.

Produces text, including transcripts, translations, language-identification responses, or answers about audio. Transcription can include punctuation and capitalization, preserve or omit fillers and repetitions, and include speaker tags. Word timestamps are optional. The supplied material does not specify a single fixed output schema.

The README provides a vLLM serving example for one H100 80 GB using vLLM 0.32.0 and the latest NeMo Speech main branch. After installing those dependencies and obtaining the checkpoint, a minimal request is:

import base64 from openai import OpenAI client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY") audio = base64.b64encode(open("audio.wav", "rb").read()).decode() response = client.chat.completions.create( model="nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B", messages=[{"role": "user", "content": [ {"type": "text", "text": "Transcribe the following audio."}, {"type": "audio_url", "audio_url": {"url": f"data:audio/wav;base64,{audio}"}}, ]}], temperature=0, extra_body={"chat_template_kwargs": {"enable_thinking": False}, "skip_special_tokens": False}, ) print(response.choices.message.content) Frequently asked questions Q: Does this model support real-time streaming?

A: No. The README describes it as an offline model and explicitly says internal audio chunking does not make it a real-time streaming ASR service.

A: The model summary states support for recordings up to one hour, with a context length of 65,536 tokens.

A: The README states that the model supports 40 languages. It refers readers to a language support matrix, but the supplied excerpt does not include that matrix.

A: Yes. Instructions can request verbatim wording or readable text with fillers and accidental repetitions omitted, and can specify punctuation, capitalization, and context terms.

Q: Can it identify speakers and provide timestamps?

A: Speaker-attributed transcription and optional word timestamps are supported. The README notes that vLLM word timestamps require the V1 model runner.

Q: What hardware does the README specify for deployment?

A: It gives one H100 80 GB as the deployment configuration for this checkpoint. The supplied material does not establish that this is the only possible hardware setup.

A: This model listing is the BF16 reference checkpoint. FP8 and NVFP4 variants are also available as separate model versions.

This is a simplified guide to an AI model called NVIDIA-Nemotron-3.5-Transcribe-30B-A4B maintained by nvidia. If you like these kinds of analysis, join AIModels.fyi or follow us on Twitter.

Nanbeige4.3-33B Explained: A 33B AI Model for Coding and Agents