Nvidia has released Nemotron 3.5 Transcribe, a free, open weight artificial intelligence model designed to convert speech into text across up to 40 languages.

Announced on Oct 28, the model is intended to help developers build applications that can transcribe audio recordings accurately.

Nvidia also introduced the second version of Granary, its open source speech dataset, to support the development and improvement of speech recognition technologies.

According to Nvidia, Nemotron 3.5 Transcribe can translate 31 languages to and from English, process recordings lasting up to one hour and distinguish between as many as eight speakers.

The company claims the model performs competitively against other speech recognition systems on industry benchmarks, including the Open ASR Leaderboard.

These capabilities could make the technology useful for businesses that need reliable transcription and translation services across different languages.

The model is designed primarily for applications such as business meeting transcription, medical documentation and other workplace services that depend on accurate speech recognition.

Developers can combine the model with Granary and Nvidia’s NeMo Speech tools to customise applications for specific industries and their technical terminology.

However, the current version processes recorded audio rather than providing real time transcription of live conversations.

Despite containing 30 billion parameters, Nvidia says the model can operate on a single chip across different hardware platforms. These include the Jetson AGX Thor robotics controller, DGX Spark computing system and Blackwell data centre chips.

The release adds to Nvidia’s collection of open weight speech models, including Parakeet, which has approximately 1.1 billion parameters and is designed to deliver efficient speech recognition.

Nvidia’s latest release comes amid growing competition in the voice AI market.

Mistral recently introduced Voxtral 3, ElevenLabs launched Realtime for voice agents that respond instantly to speech, and Salesforce released Launa Omni, its voice model.

The increasing number of releases reflects a broader industry effort to make voice based artificial intelligence more accessible, efficient and useful for businesses and consumers

Source: TechCrunch