Nemotron 3.5 Transcribe supports 40 languages and recognizes up to eight speakers

NVIDIA unveils Nemotron 3.5 Transcribe, an open voice model capable of transcribing 40 languages, identifying speakers, and translating meetings.

Transcribe recordings in 40 languages, distinguish up to eight speakers, translate conversations, and answer questions about what was said. With Nemotron 3.5 Transcribe, NVIDIA has introduced a speech recognition model designed to go beyond converting audio into text. Its goal is to combine several tasks traditionally handled by separate systems within a single multimodal architecture controlled through natural-language instructions.

Released on October 28, 2026, the model is available with open weights through Hugging Face and ModelScope. It uses a mixture-of-experts architecture with approximately 34 billion total parameters and 4.3 billion active parameters. NVIDIA is positioning it for meetings, customer-service recordings, medical transcription, and long-form conversational analysis. Transcription controlled through natural-language instructions

Traditional automatic speech recognition systems generally follow predefined formatting conventions. Some preserve hesitations, while others remove them. Numbers may appear as words or digits, and specialized terminology can be difficult to render correctly.

Nemotron 3.5 Transcribe offers a more configurable approach.

Developers can request verbatim transcription, remove disfluencies, identify speakers, or translate a recording into another language.

The same audio file can produce different outputs for different applications, including meeting minutes, subtitles, administrative records, and structured transcripts for automated analysis.

The model can also answer questions about the audio, such as identifying decisions reached during a meeting.

NVIDIA recommends disabling its reasoning mode for conventional transcription and enabling it for certain audio-understanding tasks requiring more complex interpretation. Forty languages, with different coverage for each task

NVIDIA advertises transcription in 40 languages, although the model's other functions have narrower coverage.

Capability · Supported scope Speech transcription · 40 languages Speech translation · 31 languages, to or from English Speaker-attributed transcription · 13 languages Speaker labeling · Up to 8 speakers Recording duration · Up to 60 minutes Context window · 65,536 tokens

The system can also provide word-level timestamps, allowing each part of a transcript to be aligned with its corresponding moment in the recording.

This is useful for subtitles, searchable audiovisual archives, and tools that allow users to navigate directly to specific parts of a meeting.

NVIDIA demonstrates the model using a scene from the 1940 film His Girl Friday, known for rapid dialogue and frequent overlapping speech.

The system attempts to identify individual speakers even as they interrupt one another.

Overlapping speech remains a difficult recognition problem, however. Performance can deteriorate with crowded conversations, distant microphones, or heavily compressed recordings. Identifying speakers within the same model

Speaker diarization is the process of determining who spoke and when.

Many transcription systems separate this task from speech recognition. One model transcribes the words, while another identifies speakers and assigns labels to sections of audio.

NVIDIA uses a more integrated approach.

Nemotron 3.5 Transcribe combines a Transformer speech encoder, speaker representations from Nemotron 3 Diarization, and a hybrid Mamba–Transformer language model.

Acoustic features and speaker cues are supplied to the language model, allowing it to produce transcripts with speaker labels rather than relying entirely on a separate post-processing step.

The objective is to improve consistency between recognized words and speaker assignments, especially when several people speak at once.

These labels distinguish speakers within a recording. They are not intended to establish anyone's real-world identity. Strong leaderboard results, with important differences between tests

Nemotron 3.5 Transcribe performs strongly across several public evaluations, although the reported metrics measure different capabilities.

On the English Hugging Face Open ASR Leaderboard, NVIDIA initially highlighted a 5.33% average Word Error Rate, placing first among open-weight systems in that leaderboard configuration.

The company's research report and model card also report 4.71% WER using a nine-dataset evaluation. Those figures reflect different evaluation configurations or snapshots and should not be treated as directly interchangeable.

On the multilingual Open ASR evaluation, the model scores 4.62% average WER across seven languages.

WER measures how many words are substituted, deleted, or inserted relative to a reference transcript. Lower scores indicate better recognition accuracy.

On Artificial Analysis, Nemotron 3.5 Transcribe scores approximately 2.8% WER, close to Mistral's Voxtral Small.

The comparison is notable because NVIDIA's model activates roughly 4.3 billion parameters, compared with 24 billion for Voxtral Small.

That does not make Nemotron the universally best transcription model. Rankings vary by dataset, language, scoring method, and deployment conditions. Some proprietary systems achieve better overall recognition accuracy, while other models prioritize speed or specialized tasks. Tracking overlapping speakers in difficult recordings

Speaker attribution is evaluated separately from conventional transcription.

On David AI's multilingual diarization benchmark, NVIDIA reports a Diarization Error Rate of 18.26% across 21 languages, compared with 18.38% for its specialized Nemotron 3 Diarization model.

On VoiceArena, which compares 14 systems on complete English-language sessions containing overlapping speech, Nemotron 3.5 Transcribe achieves 32.4% concatenated minimum-permutation WER, or cpWER.

The second-ranked system scores 34.2%.

This metric counts transcription errors after matching recognized words with their corresponding speakers.

A 32.4% error rate demonstrates how challenging overlapping speech remains, even for the strongest system