IBM Granite Speech 4.1 2B: Two Open Multilingual ASR Models — Autoregressive + NAR with RTFx 1,820 on H100, Apache 2.0

MarkTechPost·low signal

IBM released two compact speech models under Apache 2.0: Granite Speech 4.1 2B (autoregressive, supports ASR + bidirectional translation in EN/FR/DE/ES/PT/JA) scoring WER 1.33 on LibriSpeech clean, and a non-autoregressive variant delivering RTFx ~1,820 on a single H100 GPU by trading Japanese/translation/biasing for raw speed. A Speech-Plus variant adds speaker-attributed ASR and word-level timestamps.