NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt

NVIDIA Nemotron 3.5 Transcribe is an open-weight, promptable model that integrates transcription, speaker attribution (diarization), translation, and audio question answering into a single checkpoint, offering significant advancements in multilingual speech processing.

Core Capabilities and Performance Benchmarks

Nemotron 3.5 Transcribe is a unified speech model capable of handling complex audio tasks across multiple languages and scenarios. It supports transcription for 40 languages, translation for 31 languages (to or from English), and speaker attribution for up to 13 languages, handling recordings up to one hour long and labeling up to eight speakers.

Leaderboard Rankings

The model demonstrates top-tier performance across several industry benchmarks:

  • Hugging Face Open ASR Leaderboard: Nemotron 3.5 Transcribe ranks first among open-weight models for English transcription with an average Word Error Rate (WER) of 5.33%. It also ranks lowest among evaluated open-weight models in the Multilingual (4.62% mean WER) and Long-form (10.98% mean WER) tracks.
  • Artificial Analysis Speech-to-Text Leaderboard: It ties for first among open-weight models with a 2.8% AA-WER, achieving this accuracy with only about 4 billion active parameters, placing it on the accuracy-versus-active-parameter Pareto frontier.
  • David AI DAI-ASR-I18N Leaderboard: Transcribe ranks first among evaluated open-weight models in the combined transcription-and-speaker-attribution track and achieves the lowest diarization error rate (18.26%) across 21 languages.
  • VoiceArena Diarization-Bench: It leads among 14 systems in English full-session benchmarks, posting the lowest concatenated minimum-permutation WER (cpWER) of 32.4% in both in-person and online settings.

Natural Language Prompting for Versatile Output

Nemotron 3.5 Transcribe allows developers to guide the model's output using natural language instructions, enabling a single checkpoint to serve multiple application needs.

Task-Specific Instructions

Depending on the prompt, the same recording can be processed to achieve different results:

  • Transcription and Diarization: A prompt like “Transcribe this recording and label each speaker” generates a fully attributed transcript.
  • Translation: A prompt such as “Translate this audio into English” performs the translation.
  • Audio Question Answering (QA): A prompt like “What decisions were made in this meeting?” allows the model to extract specific information from the audio content.

Contextual Biasing

For specialized domains, developers can use context before training by providing a list of possible terms. For instance, in a healthcare setting, a prompt might read: “Transcribe this medical discussion in the original language. Possible terms: HbA1c, metformin. Use these terms only when supported by the audio. Preserve standard medical spelling and write numbers in conventional written form.”

Sources