We’re excited to share Granite 4.0 1B Speech, the newest addition to IBM’s Granite Speech collection. Designed for enterprise applications on resource-constrained devices, Granite 4.0 1B Speech is a compact speech-language model built for multilingual automatic speech recognition (ASR) and bidirectional speech translation (AST). With only half the parameters of its predecessor, granite-speech-3.3-2b, the model delivers higher English transcription accuracy, faster inference through speculative decoding, and expanded language support, now covering English, French, German, Spanish, Portuguese, and Japanese. Two recent additions on this release are Japanese ASR support and keyword list biasing for improved recognition of names and acronyms, each of that are steadily requested capabilities from the community. Granite 4.0 1B Speech also recently ranked #1 on the OpenASR leaderboard, highlighting its strong performance amongst open speech recognition systems.
Despite its small size, Granite 4.0 1B Speech achieves highly competitive results on standard English ASR benchmarks. Performance is measured using Word Error Rate (WER)—the share of words transcribed incorrectly—where lower scores indicate higher accuracy. As shown in Chart 1, Granite 4.0 1B Speech delivers strong WER across multiple datasets while using far fewer parameters than many comparable models.
Chart 1: granite-4.0-1B-speech delivers competitively low WER, thus strong ASR accuracy, across many benchmarks while being a small model.
Like all Granite models, Granite 4.0 1B Speech is released under an Apache 2.0 license with native support in transformers and vLLM. We evaluated the model across a variety of ordinary ASR and AST benchmarks — spanning English, multilingual, and translation tasks — and located it performs in addition to or higher than models well above its parameter count. Full evaluation results, architecture details, training data, and usage examples could be found on the model card. We recommend pairing it with Granite Guardian for production deployments requiring additional risk detection.
Give it a try today and tell us what you’re thinking that!
