NVIDIA on Wednesday, Oct. 28, released Nemotron 3.5 Transcribe, an AI model that turns recorded speech into text. It is an open-weight model: the model's trained files are published, so anyone can download it and run it on their own computers, and NVIDIA says it is ready for commercial use. NVIDIA says one model now does jobs that used to need several: it writes down what was said, labels who said it, translates, and answers questions about a recording.
NVIDIA's starting point is that there is no single right transcript. A multilingual call, a fast-paced meeting and a doctor's dictation each need something different, such as numbers written as digits or as words, every "uh" kept or cleaned out, or medical terms spelled correctly. NVIDIA says older speech-recognition systems settle on one style when they are trained. With Transcribe, developers ask for the style in everyday language with each recording, such as "Transcribe this recording and label each speaker" or "Translate this audio into English," and can give it a list of names or terms to expect.
It writes down speech in 40 languages, translates 31 of them to or from English, and labels speakers in 13, NVIDIA says. It handles recordings up to an hour long and can tell apart up to eight speakers. The model is free to download from Hugging Face, a site where AI developers share models, and Baseten and DeepInfra offer it as a hosted service for those who don't want to run it themselves. It works on finished recordings, not live audio, NVIDIA says.
NVIDIA says Transcribe has the best English score among models anyone can download on the Open ASR Leaderboard, a public ranking on Hugging Face of how accurately systems transcribe speech. NVIDIA's comparison leaves out paid online services. NVIDIA also says the model tops two rankings, run by other companies, of how well systems tell who spoke when. Both companies have ties to NVIDIA: it lists VoiceArena, which runs one, among its launch partners, and it notes that David AI, which runs the other, licensed NVIDIA some of the model's training data, though not the recordings its ranking tests on.
NVIDIA also says the model can be trained further on a field's own vocabulary. In its test on recordings of Indian doctors and medical students, that extra training nearly halved the share of words it got wrong, from 22.44% to 12.17%, for speakers it had not heard before, compared with the best it did from a well-worded instruction alone. Answering questions about a recording was tested only in English, and NVIDIA says it may not work reliably in other languages. NVIDIA advises checking names, speaker changes and answers against the recording, and others have yet to show how well it does on everyday recordings outside the tests NVIDIA reports.