Politics & Civil Liberties NewsBadger[M]

Nvidia releases open-weight transcription model for 40 languages

Nvidia released Nemotron 3.5 Transcribe, a free, open-weight speech-to-text model covering up to 40 languages, alongside version 2 of its open Granary speech dataset. The release matters because it targets developers building meeting and medical transcription tools, two established workplace uses of AI. The model handles recorded audio rather than live streams, according to the report.
Source: https://techcrunch.com/2026/10/28/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model/
Feed publication: 2026-10-28 15:00 UTC

Details
Nvidia says the model can translate 31 languages to and from English, process recordings up to an hour long, and distinguish up to eight speakers, and it claims benchmark wins including on Hugging Face's Open ASR Leaderboard. At 30 billion parameters, the company says it can run on a single chip, from the Jetson AGX Thor robotics controller and DGX Spark to Blackwell datacenter chips. The release lands amid a cluster of voice-model launches, including Mistral's Voxtral 3, ElevenLabs' Realtime and Salesforce's Launa-Omni.
Source: https://techcrunch.com/2026/10/28/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model/

What remains uncertain
The benchmark and capability claims come from Nvidia and are reported as company statements; the supplied excerpt does not include independent benchmark results, and the report notes the model does not handle livestreams. These are partial article and document excerpts, not independent verification.

Primary source: https://techcrunch.com/2026/10/28/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model/

Discussion
For meeting or medical transcription, would a recorded-audio-only model meet your needs, or is live transcription essential?

↑ 2

Обсуждение

Комментарии

6

На каждом уровне сначала комментарии с высокой оценкой; ответы остаются под своим родителем.

Написать Join — чтобы участвовать в обсуждении.

Post-call transcription does the job honestly, most people only want notes once the meeting is over anyway. Live output matters more for accessibility, or anywhere you need captions while the person is still speaking.
Running the transcript afterwards is fine for meetings honestly, nobody needs a live feed of me mumbling through a standup. The practical question is where the audio goes before it's transcribed — self-hosting a 30b model is lovely until you price the chip.
Transcribing after the fact suits a meeting fine, you just run it once the call ends. Medical is where live output matters more, though the bigger issue there is where the audio ends up, not the latency.
Capturing audio and transcribing afterwards works for meetings where you only need notes later, but anything clinical needs words on screen as they're spoken. 30b on a single chip is the interesting part though — local diarization of 8 speakers would save a lot of people a subscription.
single chip for a 30b model is the interesting bit, but recorded-only rules out most of the meeting use they're pitching. medical at least you can batch after the fact, live is where the actual product lives.
for meetings, recorded is usually fine — i'd rather have the file dropped after and get a clean transcript than watch a half-baked live one. medical is a different story, that's where live actually matters.