Granite Speech

IBM approaches ASR by grafting ears onto its Granite LLMs: Granite Speech 4.1 2B pairs a 16-block Conformer encoder trained with CTC and a window query transformer projector with a Granite 4.0 1B base model carrying 128K context, so one checkpoint both transcribes speech and translates it across its language set. Coverage spans English, French, German, Spanish, Portuguese, and Japanese, and the April 2026 release took a top spot on the Hugging Face Open ASR Leaderboard with a 5.33 mean word error rate at a 231 real-time factor, unusually strong numbers for a 2B model. Everything ships under Apache 2.0 with no usage restrictions, consistent with the wider Granite family's enterprise positioning. Inference paths include transformers 4.52.1 or newer, vLLM for serving, and llama.cpp and mlx-audio for local and Apple silicon use, so the model runs on modest hardware with no hosted API involved. For teams that want permissively licensed multilingual transcription with an LLM-style prompt interface, it is one of the cleanest current options.