Nvidia has introduced Nemotron 3.5 Transcribe, an AI model that developers can download for free and customize to their needs. It converts audio to text, translates speech, and can distinguish between up to eight speakers.
Nvidia has introduced Nemotron 3.5 Transcribe, an AI model that developers can download for free and customize to their needs. It converts audio to text, translates speech, and can distinguish between up to eight speakers.
This is reported by Nvidia.
The model is primarily aimed at developers of audio services. For example, it can be used to create an application that automatically records the content of business meetings, helps doctors create medical records, or converts interviews into text.
You can tell the AI assistant what kind of output you want. For example, ask it to record a conversation verbatim, including pauses and spam, or to prepare an edited text without unnecessary repetitions. You can also tell the model people's names or professional terms so that it makes fewer mistakes during recognition.
Also, the model works with ready-made audio recordings, not live conversations. Therefore, it cannot be used, for example, to instantly create subtitles during a video call.
Nemotron 3.5 Transcribe supports translation between English and 30 other languages, including Ukrainian. However, the voice recognition feature is only available for 13 languages. Ukrainian is not among them yet.
A single compatible Nvidia GPU is sufficient to run, although hardware requirements vary by model version.
Along with the model, Nvidia released Granary v2, an updated set of audio and text recordings for training the AI. With its help, and with NeMo Speech tools, developers will be able to further train the model, for example, to better recognize medical terminology, professional slang, or pronunciation features.
According to Nvidia, the new product took first place among open-scale models in the Hugging Face Open ASR Leaderboard rating for English speech recognition accuracy.
The model is already available on Hugging Face, including for commercial use under the terms of the license.
Previously, dev.ua wrote that Microsoft introduced MAI-Transcribe-2-Streaming, an AI model that can convert speech to text in real time, as well as two new voice models for AI agents.
