IBM's New Granite Speech Model Packs More Power Into Half the Size

IBM has unveiled Granite 4.0 1B Speech, a significant upgrade to its line of enterprise-grade speech models. Engineered specifically for devices with limited computing power - think handheld scanners, field equipment, or edge servers - this compact model excels at converting speech to text and translating between languages in real time.

The new version operates with roughly half the parameters of its predecessor yet improves upon it in key areas. It demonstrates higher accuracy in English transcription, processes requests faster, and broadens its multilingual support to include English, French, German, Spanish, Portuguese, and now Japanese. Two frequently requested features, Japanese speech recognition and a keyword biasing tool to better catch proper names and technical terms, are also included.

Its technical prowess is underscored by a recent #1 ranking on the OpenASR leaderboard for open speech recognition systems. In performance benchmarks, measured by Word Error Rate, Granite 4.0 1B Speech matches or surpasses the accuracy of much larger models, a notable achievement for software designed to run on constrained hardware.

Available under a permissive Apache 2.0 license, the model is ready for integration with popular frameworks. IBM provides a full model card with technical evaluations and suggests pairing it with their Granite Guardian tool for deployments where risk monitoring is essential. This release points to a clear industry trend: doing more with less, directly where the data is created.

Source: Hugging Face Blog