Nvidia released Nemotron 3.5 Transcribe, an open-weights speech language model with 30B parameters and 4.3B active parameters. The model transcribes 40 languages, adds speaker labels and word timestamps, and achieved a 4.71 percent mean WER on the English Open ASR Leaderboard (MarkTechPost).
IBM released Granite Switch 4.1.1, an approach that trains an external instruction layer on task feedback while leaving the base model and harness unchanged, enabling model updates without full retraining (Hugging Face Blog). Separately, Block’s AI lab released Dessa, a full-duplex speech-to-speech model trained only on human conversation audio that runs locally on a MacBook Pro M4 Max, though Block has published no model weights or detailed benchmarks (MarkTechPost).