NVIDIA’s Nemotron 3.5 Transcribe ties for #1 among open-weight speech-to-text models at 2.8% WER

What happened

Artificial Analysis reports that NVIDIA’s new Nemotron 3.5 Transcribe ties Mistral’s Voxtral Small for #1 among open-weight models on its AA-WER Index at 2.8% WER. The non-streaming transcription model has 30B total and 4B active parameters, matching Voxtral Small’s accuracy with 6x fewer active parameters (4B vs. 24B). On AA-AgentTalk, Artificial Analysis’s voice agent conversation dataset, it scores 1.8% WER versus Voxtral Small’s 2.3%.

Read next

• More changes in today’s briefingGoogle DeepMind announces Genie 4 for real-time world simulation with native audio

The story so far

Artificial Analysis ranked NVIDIA’s Nemotron 3.5 Transcribe joint first among open-weight speech-to-text models on its AA-WER Index, tying Mistral’s Voxtral Small with a 2.8% word error rate. Nemotron 3.5 Transcribe is a non-streaming transcription model with 30B total parameters and 4B active parameters, compared with Voxtral Small’s 24B active parameters. On AA-AgentTalk, Artificial Analysis’s dataset of voice agent conversations, the evaluator reported a 1.8% word error rate for Nemotron 3.5 Transcribe versus 2.3% for Voxtral Small.