Nvidia releases a speech model covering 40 languages
Nvidia has released Nemotron 3.5 Transcribe with downloadable weights for speech transcription, speaker attribution and translation. The model covers 40 languages and arrives alongside Granary v2 training data and adaptation recipes. Developers can run the system within their own infrastructure and tune it for particular domains, rather than relying solely on a hosted transcription service.
Artificial Intelligence··Night
One speech model covers several tasks
Nvidia, the chipmaker and AI developer, released Nemotron 3.5 Transcribe on 28 October with downloadable model weights. The speech system transcribes recordings in 40 languages and combines transcription with speaker attribution, translation and questions about audio. The release gives developers a checkpoint they can operate on their own infrastructure, alongside the Granary v2 speech data and tools for adapting recognition to particular uses.[1], [2]
Natural-language instructions guide the output. A meeting recording can become a transcript with speakers identified, a translation, or an answer about decisions made during the discussion. Nvidia says the system can process recordings of up to an hour and distinguish up to eight speakers. It is designed for previously recorded audio; the release does not provide a live-streaming transcription system.[1]
Language support depends on the requested task
Language coverage in Nvidia’s model varies by function. Transcription supports 40 languages, translation works between English and 31 languages, and speaker attribution covers 13. Audio question answering was trained in English. Developers can also provide likely specialist terms in their instructions so the model has context for vocabulary it encounters, with the requested terms used when the recording supports them.[1]
The model also supplies timestamps for individual words, allowing a transcript to be aligned with playback. Nvidia describes a classroom demonstration by Agora, a company supplying real-time audio and video infrastructure, that places the transcript beside a replay. The released tools address applications such as meeting software and medical dictation, where recognizing the words and identifying their speaker are both useful parts of the output.[1]
Speech data supports further adaptation
Granary v2, Nvidia’s released speech-data collection and curation pipeline, supplies aligned transcript forms and information about data quality. Developers can choose written-output conventions and adapt the model with NeMo Speech, Nvidia’s speech-training toolkit. The materials include guidance for preparing domain vocabulary and evaluating recognition. The accompanying workflow links representative audio with the written form an application needs, so adaptation can address specialized language as well as ordinary transcription.[1]
