NVIDIA shipped Nemotron 3.5 Transcribe as open weights: one 30B model that transcribes 40 languages, translates 31, labels up to eight speakers and answers questions about the audio
AspireAI community · NVIDIA on Hugging Face
NVIDIA released Nemotron 3.5 Transcribe today as an open-weight speech model that folds four jobs into one checkpoint: transcription, speaker attribution, translation and audio question answering. It transcribes 40 languages, translates 31 of them to or from English, attributes speakers in 13, handles recordings up to an hour, and labels up to eight speakers. Architecture is a hybrid Mamba-Transformer mixture-of-experts language model with roughly 4.3B active parameters out of 30B, shipped in BF16, FP8 and NVFP4.
The results are strong for an open model. It is first among open-weight models on the Hugging Face Open ASR Leaderboard for English, which the launch blog reports at 5.33 percent average word error rate with API models excluded, while the model card gives 4.71 percent mean WER over the nine public sets; it is lowest among open-weight models on the multilingual track at 4.62 percent and the long-form track at 10.98 percent, and ties for first on Artificial Analysis at 2.8 percent AA-WER with about 4B active parameters. On speaker attribution it posts the lowest diarization error rate of 14 systems tested on David AI's 21-language benchmark at 18.26 percent, and leads VoiceArena's 14-system full-session comparison.
The honest limitation: NVIDIA's own fine-tuning case study reports a 45.8 percent relative WER cut on held-out speakers, but that is a single medical domain and the benchmark rankings are the vendor's own submissions, not an independent third-party sweep. NVIDIA also flags that language support differs by task and that audio question answering is trained in English only, and the model is built for recorded audio rather than live streaming. Quantized versions and local deployment are stated as coming soon.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.