A 4B-active transcription model: can it replace Whisper plus speaker diarization?
NVIDIA has packed transcription, speaker attribution, translation, and audio Q&A into one open-weights model. The leaderboard looks good, but real meeting rooms still have overlapping speech, mixed languages, and a VRAM bill to get through.
NVIDIA released Nemotron 3.5 Transcribe, an open-weights 30B MoE model with about 4B active parameters. It aims to handle transcription, speaker attribution, translation, contextual prompting, and audio understanding with the same checkpoint.
The official demo uses the fast, overlapping dialogue from the 1940 film His Girl Friday as an example. The choice is straightforward: the most annoying part of meeting notes is often not failing to hear the words, but not knowing who said them. Transcription, speaker labels, translation, and word-level timestamps are all packed into a single generation, eliminating the pipeline that stitches together Whisper, diarization, and post-processing.
For meeting recordings, it can label different speakers as spk1 and spk2, and it can also translate into Spanish or Japanese on prompt, or find PII. The NVIDIA blog says it transcribes 40 languages, translates between 31 languages and English, and supports speaker attribution for 13 languages. It handles recordings up to 1 hour long, up to 8 speakers, and a 65,536-token context.
The benchmark results are also extensive. It ranks first in VoiceArena's 14-system diarization plus ASR comparison, and it also ranks first on David AI's 21-language diarization leaderboard. The official blog says it is first among open-source English models on Hugging Face Open ASR, with an average WER of 5.33%. The model card later lists 4.71%; the two figures use different methodologies, so don't compare them directly.
The Artificial Analysis data is more interesting. Nemotron 3.5 Transcribe gets 2.81% AA-WER, while Voxtral Small gets 2.77%; the latter has 24B active parameters. Accuracy is almost tied, while active parameters differ by about six times. This is also the key point of the MoE narrative: each token only goes through a small subset of parameters.
But 4B active does not mean 4B VRAM. A commenter in the replies pointed out that with 4-bit quantization, 30B weights still take about 15GB; once audio context is included, a 24GB GPU or 32GB Mac is more realistic, and ordinary office machines should not get too excited. NVIDIA says the quantized version will support RTX Spark, DGX Spark, and Jetson AGX Thor, and can also run on Blackwell data-center GPUs. Note that this is an official plan, not something already running on your old laptop.
The scope of the open-source release is also worth examining. The weights, Granary v2 data, and fine-tuning recipe are all released. The public Granary v2 manifest lists 463,062 hours, 48 languages, and 16 corpora. In a medical ASR case, fine-tuning reduced WER on held-out speakers from 22.44% to 12.17%, a relative reduction of 45.8%. Results like this are closer to real business than a No. 1 leaderboard spot, because departmental accents, drug names, and abbreviations don't clean themselves up.
The scenarios that really matter are very concrete. Four people in a meeting all saying "sorry, go ahead" at the same time, two people interrupting each other, Chinese mixed with English, Urdu mixed with English, and rickshaw horns. Some commenters directly asked how it performs on voice notes from Karachi, while others wanted to see an Italian family dinner. These are not things a leaderboard can stand in for. The model card also notes that speaker labels only distinguish different people in the recording; they do not equal real identities. It is an offline model, not real-time streaming ASR. Audio understanding and reasoning were evaluated only in English; don't assume they are available in other languages by default.
If someone runs it on a real customer meeting next week and the transcript gets names, numbers, and who made the decision right, that will count as a win. A No. 1 leaderboard spot is only a ticket to entry.
Original: https://x.com/testingcatalog/status/2115697759281975475
Model: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
Demo: https://huggingface.co/spaces/nvidia/nemotron-transcribe
Blog: https://huggingface.co/blog/nvidia/nemotron-transcribe-granary
发布时间: 2026-10-29 16:32