Nvidia releases Nemotron 3.5 Transcribe for multilingual speech-to-text

A new tool for speech processing

Nvidia has introduced Nemotron 3.5 Transcribe, a new AI model designed to simplify the process of converting spoken audio into written text. The model is built to handle various tasks that are common in business environments, such as multilingual transcription and translation. By providing this tool, Nvidia aims to help developers create more efficient ways to process recorded conversations.

The model is unique because it is released with open weights. This means that developers can access the underlying trained parameters of the model to build their own specific applications. Instead of being forced to use a closed third-party service, companies can host and customize the technology to fit their own needs. This provides a level of flexibility that is often missing in standard commercial transcription software.

As reported by Convergence Now, the model is designed to be highly capable across many different linguistic barriers. It can translate between English and 31 other languages while also transcribing speech in up to 40 different languages. This broad support makes it a versatile option for global organizations that need to manage communications across various regions and dialects.

Technical capabilities and limits

One of the standout features of Nemotron 3.5 Transcribe is its ability to handle complex audio environments. The model can distinguish between as many as eight different speakers in a single recording. This is a significant capability for developers who need to create transcripts of group meetings or multi-person discussions where identifying who said what is essential. It can process recordings that are up to one hour long, making it suitable for substantial meetings.

To help developers improve the performance of these models, Nvidia also released Granary v2. This is an updated open-source speech dataset that provides the raw data needed to train and refine voice AI applications. By combining the model with this dataset and Nvidia's NeMo Speech tools, developers can fine-tune the AI for specific use cases. This is especially helpful when general-purpose systems fail to recognize technical vocabulary or industry-specific expressions.

Nvidia claims that the model performs competitively on several industry benchmarks, including the Hugging Face Open ASR Leaderboard. ASR stands for Automatic Speech Recognition, which is the technology used to turn spoken words into text. While these performance claims come directly from the company, they serve as a starting point for developers who will need to conduct their own independent evaluations to see how the model handles their specific data.

Applications in business and health

The release of this model has immediate implications for various business sectors, particularly those that generate large amounts of spoken data. For example, companies can use the tool to turn recorded meetings into written transcripts. This allows employees to review discussions more easily, identify key decisions, and search for specific information within a text format rather than listening to hours of audio.

Healthcare is another primary area where this technology could be applied. Medical professionals frequently use specialized terminology, including complex drug names and clinical language that standard AI might misinterpret. A transcription model that can be adapted to recognize these specific terms could help developers build tools for creating clinical documentation. This could streamline the way doctors and nurses record patient interactions and medical notes.

However, the use of AI in high-stakes environments like medicine requires caution. Because accuracy is so important in clinical records, any text generated by the model must be verified by a human professional. The goal for developers in this space is to create a tool that reduces the manual workload of transcription while maintaining the high level of precision required for patient safety.

The shift toward open weights

By releasing Nemotron 3.5 Transcribe with open weights, Nvidia is participating in a broader trend toward open-source AI models. This approach allows for greater customization than many