Tool Suggestions
NVIDIA Nemotron 3.5 Transcribe: One Free Open Model for Transcripts, Speakers and Translation
NVIDIA released Nemotron 3.5 Transcribe on October 28: a free, open-weight speech model that folds four jobs into one checkpoint — transcription, speaker labels, translation and questions about the audio — steered by plain-language prompts. It runs on a single GPU.
What one checkpoint does
Most speech tools chain separate steps: detect speech, write it down, label speakers, summarize. Transcribe does it in a single pass. One recording can produce different outputs depending on the prompt: a verbatim transcript, a cleaned-up written form, speaker-attributed text, a translation, or answers about what was decided in a meeting. You can also hand it jargon and name lists so specialized vocabulary is spelled correctly — useful for medical and financial recordings.
The specs
A 30-billion-parameter mixture-of-experts model with about 4 billion active parameters per token. It transcribes 40 languages, translates 31 of them to and from English, labels up to 8 speakers across 13 languages, handles recordings up to an hour long, and returns word-level timestamps. It ships with Granary v2, NVIDIA’s bigger open speech dataset and pipeline for teaching it your own vocabulary. For hardware: FP8 (about 35 GB) and NVFP4 (about 21 GB) checkpoints exist — the NVFP4 build runs on an RTX 5090 — and NVIDIA says one RTX PRO 6000 Blackwell Server Edition GPU processed 100 hours of audio in under three minutes (measured on 30-second clips at high concurrency, so real results will vary). It runs from Jetson AGX Thor edge chips and DGX Spark desktops up to Blackwell datacenter cards.
How it scores — and where it falls short
NVIDIA says it ties Mistral’s Voxtral Small for first place among open-weight models on the Artificial Analysis speech-to-text index (2.8% AA-WER), leads open models on the English Open ASR Leaderboard, and posts the lowest diarization error of 14 systems on a 21-language leaderboard. Those are NVIDIA’s claims from its own report. The honest limits: it works on recorded audio, not livestreams — if you need real-time captions, this is not the tool. Audio question-answering is trained in English only. Far-field Mandarin meetings are a weak spot (one independent analysis measured 34.95% cpCER on the AliMeeting set). And top commercial transcription APIs still lead on accuracy.
Who should care
Developers building meeting and medical transcription apps, and anyone who wants transcription without sending audio to a third-party API. It joins a busy stretch of voice releases — Mistral’s Voxtral 3, ElevenLabs Realtime, Salesforce’s Launa-Omni — and it is the most capable open option for turning recordings into usable text on your own hardware.