NVIDIA Speech Transcription Tops Open-Source Leaderboard: 4B Activated Parameters Are Very Economical, but Don't Forget the 30B Memory Bill

NVIDIA's Nemotron 3.5 Transcribe scored 2.8% AA-WER on Artificial Analysis's non-streaming speech-to-text leaderboard, tying Mistral's Voxtral Small for first place among open-weight models. It has only 4B activated parameters and 30B total parameters, using a MoE architecture. The evaluator's comparison is direct: Voxtral Small uses 24B activated parameters, while Nemotron 3.5 Transcribe matches its WER with 6x fewer activated parameters. The previous generation's most accurate NVIDIA model, Canary Qwen 2.5B, was 4.3%, and Thinking Machines' Inkling was 3.5%.

The No. 1 spot on the leaderboard looks great, but product managers should not rush to hook it up to a real-time voice assistant. The key word is non-streaming. Nemotron 3.5 Transcribe does offline transcription; streaming recognition is a separate matter. Some commenters said non-streaming is a serious drawback for voice agents. This reminder matters more than the headline.

On Artificial Analysis's AA-AgentTalk dataset, Nemotron 3.5 Transcribe achieves 1.8% WER, ahead of Voxtral Small's 2.3%. AA-AgentTalk is voice-agent conversation data. This number looks good, but do not misread it as meaning it can be directly plugged into a real-time voice agent. Some commenters dug up NVIDIA's model card and noted that Nemotron 3.5 Transcribe is an offline model; the streaming model is a separate 0.6B Nemotron 3.5 ASR.

On speed and cost, Artificial Analysis tested on a hosted DeepInfra endpoint. Nemotron 3.5 Transcribe has a Speed Factor of 75x, transcribing 1 minute of audio in under 1 second, at $2 per 1,000 minutes of audio. For batch transcription, subtitle archiving, and meeting-recording cleanup, this combination is attractive.

But the bill must be calculated in full. 4B activated parameters do not mean the model only occupies 4B of memory locally. Some commenters noted that running local inference still requires memory for 30B total parameters; what is cheap is the computation per token, while model size is separate. Another commenter said that with 4B activated parameters matching 24B, those doing local subtitle transcription can prepare to switch models. These two statements are not contradictory: one is about compute, the other about VRAM.

Multilingual performance is also a question mark. Some commenters asked about German and Brazilian Portuguese performance and pointed out that many models are very good in English but notably worse in other languages. Some mentioned Turkish and code-switching, where developers mix English technical terms into sentences; this kind of mixed speech is the real test. Others wanted to see performance on accents, noisy phone recordings, and overlapping speech; leaderboards often diverge from reality in these scenarios.

Pricing is also disputed. $2 per 1,000 minutes is not expensive for individual batch processing, but it is another matter for a high-concurrency call center. Some commenters believe that for scaled voice-agent scenarios this price is still on the high side.

Some commenters dug up NVIDIA's model card and added: it supports transcription in 40 languages, of which 31 can be translated to and from English, 13 come with speaker labels, and it can process recordings up to 1 hour long. These capabilities are suited to offline batch processing. For real-time dictation and voice agents, look at the streaming version first; do not force-fit the No. 1 non-streaming leaderboard result onto them.

Full results are on Artificial Analysis's speech-to-text page, and the evaluation methodology is here. The original tweet is here.