Nvidia Corp. released Nemotron 3.5 Transcribe, an open-weight AI model that handles multilingual audio transcription and translation, along with an open data pipeline for fine-tuning it. The model is aimed at developers who need speech processing that runs from edge devices to data centers.
Transcribe completes in a single pass what older transcription systems split across several stages: detecting speech, transcribing it, labeling speakers and answering questions about the text. According to Nvidia, the model can ingest and answer questions about 100 hours of audio across 40 languages and can identify up to eight speakers.
The model is a 30-billion parameter mixture-of-experts design with 4 billion active parameters, small enough to fit on modest hardware while running faster on more powerful equipment. It operates on a single Nvidia GPU, ranging from the Jetson AGX Thor edge computer and the DGX Spark desktop AI supercomputer to a Blackwell data-center GPU. Nvidia said a Blackwell Server Edition GPU using NVFP4 precision with one-quarter model compression can process 100 hours of audio in under three minutes.
Nvidia reported that Transcribe achieves a 30.8% word error rate out of the box, reducible to 20.8% with careful prompting. When fine-tuned for specialized work, such as the Eka Care Indian medical speech benchmark, the model reached an 11.8% word error rate, a substantial reduction.
Users can direct the model through prompts, instructing it to transcribe and label speakers, translate audio into English, or summarize decisions made in a meeting, or combine such requests. The same checkpoint can produce different outputs with separate prompts, and users can supply jargon or dialogue in advance to improve accuracy. Nvidia built the model by extending its open-source multilingual speech dataset, Granary v1, into Granary v2 using Nvidia NeMo Curator. The public manifest for Granary v2 covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in the earlier version. Those figures describe the dataset release rather than Transcribe's full training mixture or sampling weights.
Granary v2 produces three transcripts for each recording in clean verbatim styles that can be requested through prompts: one retaining filler words such as "ums" and "ahs," one converting spelled-out numbers to figures, and one removing filler words. The pipeline is intended to let developers adapt transcription across languages into readable formats and fine-tune model outputs. Transcribe is available as a Hugging Face demo, where developers can download a checkpoint and adapt it for domain expertise.
More AI news from TechManNews.
