Nvidia has expanded its portfolio of artificial intelligence tools with Nemotron 3.5 Transcribe, a new open model designed for audio transcription, and simultaneously introduced the second version of its open Granary dataset. These innovations are designed with developers in mind who are creating advanced natural language processing applications.
Nemotron 3.5 Transcribe excels in its ability to transcribe audio recordings in up to 40 languages and offers bidirectional translation between 31 languages and English. The model can process recordings up to one hour in length and is capable of distinguishing up to eight different speakers within a single dialogue, which is crucial for applications focused on conference calls or interviews.
Haben Sie eine Idee für eine Website oder App?
Lassen Sie uns unverbindlich darüber sprechen – ohne Druck, mit einem konkreten nächsten Schritt.
In internal tests and comparisons with competitors, Nemotron 3.5 Transcribe demonstrably achieves top-tier results. It is primarily intended for developers creating applications for corporate meetings and medical transcription, areas where accurate and reliable speech transcription is essential. By combining it with the updated Granary dataset and tools like NeMo Speech, developers can customize their applications to recognize specific terminology and jargon in a given industry.
One limitation of the current version is that the model only works with pre-recorded audio and does not yet support live stream transcription. Nevertheless, it is an impressive tool with 30 billion parameters, designed to run efficiently on a single chip, from embedded devices like Jetson AGX Thor and laptop chips like DGX Spark, to powerful Blackwell data center chips. Together with other open-weight models, such as the compact but powerful Parakeet with 1.1 billion parameters, Nvidia continues to strengthen its position in open AI solutions.
Nvidia's new solution comes at a time of increased activity in the voice technology market, where there is a clear trend towards the development and release of increasingly advanced models for working with speech and voice.