Nvidia Releases Open-Weight Nemotron 3.5 Transcribe for Multilingual Audio
Nvidia has released Nemotron 3.5 Transcribe, an open-weight speech model supporting 40 languages. The 34B-parameter AI handles transcription, translation, and speaker diarization for recordings up to one hour. Nvidia also launched Granary v2, a dataset of over 463,000 hours of speech. The model ranks first among open-weight options on major leaderboards.
Key points
- Nvidia, the American multinational technology company, released Nemotron 3.5 Transcribe, an open-weight speech model capable of transcribing 40 languages and translating 31 of them to or from English.
- The model features 34 billion parameters but utilizes only approximately 4 billion active parameters per step, enabling it to run on a single Nvidia chip while handling recordings up to one hour long and distinguishing up to eight speakers.
- Nvidia simultaneously launched Granary v2, a free dataset containing 463,062 hours of speech across 48 languages, which serves as the foundation for the model's training and fine-tuning capabilities.
- The model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English with a 5.33% word error rate and holds the lowest error rate on David AI’s 21-language diarization leaderboard.
- The software is designed for recorded audio rather than live transcription, allowing developers to use natural language prompts to customize outputs for specific domains such as medical dictation or technical meetings.
- The release positions Nvidia competitively against commercial rivals like SpaceXAI’s Grok, which charges for similar transcription services, by offering a commercially usable, free alternative that supports global multilingual applications.
Nvidia Expands Open AI Portfolio with Multilingual Transcription Model
Nvidia has officially released Nemotron 3.5 Transcribe, a new open-weight artificial intelligence model designed to convert recorded speech into text across 40 different languages. The announcement, made on Wednesday, marks a significant expansion in Nvidia’s portfolio of open-source tools, providing developers with a comprehensive solution for audio processing that includes transcription, translation, speaker attribution, and audio-based question answering within a single checkpoint. The model is now available for download on Hugging Face and ModelScope, signaling Nvidia’s continued push to dominate the open-weight AI landscape.
The core capability of Nemotron 3.5 Transcribe lies in its multilingual proficiency. While it can transcribe speech in 40 languages, it can translate 31 of these languages to or from English. This functionality addresses a common pain point in global business and media, where multilingual calls, fast-paced meetings, and medical dictations require not just raw text, but structured data that identifies who said what and how to format domain-specific jargon. The model is engineered to handle recordings up to one hour in length and can attribute speakers in up to eight distinct voices, a feature known as diarization.
Technical Architecture and Performance Metrics
From a technical standpoint, the model comprises 34 billion parameters. However, Nvidia highlights that only approximately 4 billion parameters are active during each processing step. This architectural choice is designed to keep the model fast and cost-effective to run, allowing it to operate on a single Nvidia chip, ranging from small Jetson edge devices to large data-center GPUs. The license accompanying the model permits commercial use, a critical factor for enterprises looking to integrate speech recognition into their products without incurring licensing fees.
Performance benchmarks indicate strong results for an open-weight model. Nvidia states that Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English, achieving an average word error rate of 5.33%. Additionally, the model sits on the accuracy–active-parameter frontier as noted by Artificial Analysis. For speaker attribution, it posts the lowest error rate on David AI’s 21-language diarization leaderboard and ranks first in a 14-system comparison by VoiceArena. Quantized versions of the model, which further reduce resource requirements, are expected to be released soon.
Granary v2 Dataset and Developer Tools
Alongside the model release, Nvidia introduced Granary v2, a free dataset comprising 463,062 hours of speech across 48 languages. This is a substantial increase from the first version, which included 25 languages. The dataset is intended to help developers fine-tune the model to adapt recognition to specific domain vocabularies and accents. Developers can download the audio from each original source under its respective license, ensuring compliance with source material rights.
The model is designed to be promptable, allowing developers to guide output using natural language. For example, users can request a clean transcript, a word-for-word transcript, or ask the model to listen for specific medical or product terms. This flexibility makes the tool suitable for various applications, from transcribing medical dictations to formatting technical meeting notes. However, Nvidia explicitly notes that the model is built for recorded audio and is not intended for live transcription tasks.
Market Context and Competitive Landscape
The release of Nemotron 3.5 Transcribe places Nvidia in direct competition with commercial speech-to-text services. Rivals such as SpaceXAI’s Grok transcription service charge users for similar jobs, whereas Nvidia’s offering is free and open-weight. This strategy allows Nvidia to attract developers and enterprises looking to reduce costs while maintaining high accuracy. The ability to run on a single chip also lowers the barrier to entry for smaller developers who may not have access to extensive computational resources.
As Nvidia continues to open-source its models, the company aims to establish its technology as the standard for multilingual audio processing. The combination of high performance, multilingual support, and commercial usability positions Nemotron 3.5 Transcribe as a significant tool for global applications. The upcoming release of quantized versions will further expand the model’s accessibility, potentially accelerating adoption across various industries that rely on accurate and efficient speech recognition technologies.
Sources
The WireByte editorial team synthesises technology news from multiple primary sources, verifies the facts, and links every source. Articles are produced with AI assistance and reviewed under our editorial policy.
