NVIDIA Releases Speech Transcription and Translation Model Nemotron 3.5 Transcribe

On October 28, 2026, NVIDIA released "Nemotron 3.5 Transcribe," an AI model that performs speech transcription and translation. By adding text instructions to a recording, users can specify whether to keep disfluencies in the transcription or omit them for easier reading. It supports 40 languages including Japanese and can transcribe recordings up to 1 hour long. The trained model and tools for additional training are also being released.

NVIDIA Nemotron 3.5 Transcribe: Intelligent Transcription with a Speech LLM - NVIDIA Nemotron

Meet NVIDIA Nemotron 3.5 Transcribe 🎙️

An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.

Watch it keep track of who said what when everyone’s talking at once.

1/ 🧵— NVIDIA AI (@NVIDIAAI) October 28, 2026

Specify whether to keep utterances as-is or clean them up for readability

The writing style desired for transcription differs by use case. When faithfully recording utterances, one may want to keep disfluencies and repetitions, while when producing readable text, one may want to omit them. NVIDIA explains that with conventional speech recognition, the format of the output transcription is fixed at training time.

Nemotron 3.5 Transcribe combines an encoder that reads audio, a model that distinguishes speakers, and a large language model, and transcribes according to text instructions given along with the recording. It can also specify whether numbers should be written out as words or expressed as Arabic numerals. These output formats can be changed through per-request instructions, without the need to retrain the model.

It can also take candidate words likely to appear in the audio, such as personal names and technical terms, as text to aid recognition.

Supports Japanese speaker-attributed transcription and bidirectional translation with English

According to the language support table in the model card, for Japanese it supports speech recognition, multi-speaker transcription, and bidirectional translation with English. It also supports code-switching, in which multiple languages are used within a single recording. NVIDIA classifies Japanese quality as "Production-ready," equivalent to a level suitable for production use.

In multi-speaker transcription, even in recordings where utterances overlap, it can label each speaker and output at what second each word's utterance starts and ends in the recording. According to the technical report, it can distinguish up to 8 speakers.

For translation, it generates English text from Japanese speech and Japanese text from English speech. Transcription and translation can be used for post-processing recordings of meetings, calls, and the like. It does not support real-time processing that transcribes sequentially while someone is speaking.

As transcription aids, it can also answer questions about the audio content and respond to instructions given by voice. However, these have only been evaluated in English, and may not work stably in other languages including Japanese, according to the company.

Evaluating multilingual and multi-speaker speech recognition

In its technical report, NVIDIA presented multilingual evaluation results from the Open ASR Leaderboard, which compares speech recognition performance. The average word error rate (WER, lower is better) across 7 languages including German and French was 4.62%. Among the models listed in the comparison table, this was the lowest value among those with publicly released weights (numerical data adjusted through training). These 7 languages do not include Japanese.

Nemotron 3.5 Transcribe Technical Report (Table 9) - NVIDIA

Among the compared models, Qwen3-ASR-1.7B was 6.49% and Whisper-large-v3 was 7.94%. Meanwhile, the commercial model ElevenLabs Scribe v2 was 3.56%, a lower value than Nemotron 3.5 Transcribe. The figures for these three models use the ranking published values as of October 2, 2026.

For multi-speaker transcription, using recordings of English meetings, calls, and the like, error rates were compared including whether utterances were assigned to the correct speaker. The comparison targets were four systems with publicly released model weights, including ones that combine speech recognition with processing that determines who spoke when. As a result of evaluating with the same data and scoring method, the error rate of Nemotron 3.5 Transcribe was lower than the other four systems on average across 6 conditions with different microphone placements and datasets [1].

Releasing model weights and tools for additional training

The model is distributed on Hugging Face as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. It supports commercial and non-commercial use, and adopts OpenMDW-1.1 as its license.

The total number of parameters including the audio encoder is approximately 34 billion, of which approximately 4.3 billion are used during processing.

In addition to the BF16 version, quantized FP8 and NVFP4 versions that represent weights with fewer bits to reduce memory usage are available. The NVFP4 version is provided for Blackwell GPUs.

Regarding the quantized versions, NVIDIA recommends the FP8 version when recognition accuracy in languages other than English is a priority. When memory capacity or throughput per unit time is the main constraint, it points to the NVFP4 version [2].

The model card includes procedures for running the BF16 version using a single 80GB H100 with the inference engines vLLM and NeMo Speech. A GGUF format is also provided for NeMo-Speech.cpp, a C++ runtime for local execution.

In processing performance measurements, a single RTX PRO 6000 Blackwell Server Edition was used, combining the NVFP4 version with the auxiliary model DFlash2 that accelerates inference. When processing multiple 30-second recordings simultaneously with the optimized inference environment Nemotron Transcribe NIM, it reported being able to transcribe up to approximately 37 minutes (2,238 seconds) of audio per second in total. However, this optimization is not included in the open-source version at the time of release.

Tools for additional training tailored to one's own language, domain, and tasks have been released in NeMo Speech. For detailed inference procedures and supported environments, refer to the model card.