IBM introduces Granite 4.0 1B Speech for edge AI
IBM introduces Granite 4.0 1B Speech, a compact voice model for transcription and translation in six languages, including Japanese. It uses half as many parameters as its predecessor and adds keyword recognition, with a design intended to run on resource-constrained devices.
IBM has introduced Granite 4.0 1B Speech, a compact model that can transcribe and translate conversations in several languages on resource-constrained devices. It is designed for businesses that need to process voice without relying on the cloud every time.
The model has half the parameters of its predecessor, granite-speech-3.3-2b, while improving English transcription accuracy. It also delivers faster inference through speculative decoding, a technique that speeds up results by anticipating part of the response.
Six languages and voice translation
Granite 4.0 1B Speech supports automatic speech recognition, known as ASR, and bidirectional speech translation, or AST. In practice, it can turn a spoken conversation into text and translate it between languages.
The new version works with:
- English
- French
- German
- Spanish
- Portuguese
- Japanese
Japanese is one of the new features in this release. The other is keyword list biasing, a function that helps the model recognize proper names, acronyms and specific terms. For example, a company could provide a list of its brands, products or internal acronyms to reduce errors during a call.
A small model with competitive results
IBM measures accuracy using word error rate, or WER. This metric represents the percentage of words the system transcribes incorrectly: the lower the result, the better it performs.
According to IBM's evaluations, the model delivers competitive results across several standard English speech recognition tests, despite having far fewer parameters than other comparable systems. The company also says it matches or outperforms considerably larger models on various recognition and translation tasks.
In addition, Granite 4.0 1B Speech recently took first place on the OpenASR leaderboard, which compares open automatic speech recognition systems. That result points to strong performance among publicly available models, although you still need to test it with the accents, background noise and vocabulary of each use case.
What changes for you
Its smaller size allows the model to run on more modest hardware, such as industrial devices, customer service systems or applications that need to respond without sending every recording to an external server. This can reduce latency, infrastructure costs and exposure of sensitive data.
The model is released under the Apache 2.0 license and has native support in transformers and vLLM, two common tools for running AI models. IBM recommends combining it with Granite Guardian in production environments that need to detect additional risks.
The important next step is to check its performance outside benchmark tests, with background noise, different accents and real conversations. If it maintains its results under those conditions, small voice models could become a practical way to bring AI directly to the device.