IBM has announced the release of Granite 4.0 1B Speech, the latest iteration in its Granite Speech collection. This model is specifically engineered for enterprise applications on devices with limited resources, offering a compact solution for multilingual automatic speech recognition (ASR) and bidirectional speech translation (AST).
Granite 4.0 1B Speech features only half the parameters of its predecessor, granite-speech-3.3-2b, yet it achieves superior transcription accuracy in English and faster inference times through a technique known as speculative decoding. The model now supports a wider array of languages, including English, French, German, Spanish, Portuguese, and Japanese. Notably, this release introduces Japanese ASR support and keyword list biasing, enhancing the model’s ability to recognize names and acronyms—features that have been highly requested by users.
In a significant achievement, Granite 4.0 1B Speech has secured the top position on the OpenASR leaderboard, underscoring its robust performance among open speech recognition systems. Despite its compact size, it delivers competitive results on standard English ASR benchmarks, measured by Word Error Rate (WER), where lower scores signify greater accuracy. The model’s performance is illustrated in Chart 1, which demonstrates its ability to maintain low WER across various datasets while utilizing significantly fewer parameters than many of its counterparts.
As with all Granite models, Granite 4.0 1B Speech is available under an Apache 2.0 license and offers native support in transformers and vLLM. Comprehensive evaluation results, architectural details, training data, and usage examples can be found on the model card. For production deployments that necessitate additional risk detection, it is recommended to pair this model with Granite Guardian.
IBM encourages users to explore Granite 4.0 1B Speech and share their feedback.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
