Nvidia just killed paid transcription with a free AI that hears 40 languages

Nvidia has released Nemotron 3.5 Transcribe, an open model for turning recordings into text. The weights are free on Hugging Face and ModelScope, and there is a demo where you can upload your own file and see what comes out.

It handles speech in 40 languages, translates to and from English for 31 of them and can tell apart up to eight speakers, so the transcript shows who said what. A recording can run up to an hour. You steer it with plain text: feed it product names, surnames and industry jargon in advance, ask for a verbatim or a cleaned up transcript, or ask questions about what was said.

According to Nvidia, Transcribe ranks first among open models on the English Open ASR Leaderboard on Hugging Face, and in speaker labelling it posts the lowest error rate on David AI's 21 language leaderboard. It has 30 billion parameters, about 4 billion of them active at a time. You can serve it with vLLM today, and quantized versions for local hardware like DGX Spark and Jetson AGX Thor are promised later.

The catch: it only works with recorded audio, not live streams. Alongside the model Nvidia updated its open speech dataset Granary to version 2 and published fine-tuning recipes for specific fields such as medicine.

Meeting transcription services sell exactly this on a subscription. Now they have a free rival you can run yourself.