Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster

Why it matters

IBM has expanded its Granite Speech lineup with a new speech-language model and refreshed versions of two existing ones, aiming to cover the main demands of enterprise audio work. The new model, granite-speech-5.0-2b, handles transcription across six languages plus translation to and from English, and adds practical controls like language identification prompting, keyword biasing for domain-specific terms, and on-request punctuation and capitalization. It also improves robustness in noisy, reverberant, and distant-microphone settings. The revised granite-speech-5.0-470m-turboctc keeps its compact encoder design while adding multilingual transcription, native punctuation, and better handling of echoey rooms, and it runs considerably faster than at launch while maintaining accuracy. The updated granite-speech-4.1-2b-plus targets multi-speaker recordings such as meetings and support calls, improving speaker attribution, timestamp precision, and long-form transcription, and serving as a drop-in replacement since its architecture and API are unchanged. All three are available under an Apache 2.0 license with Transformers support, and IBM recommends pairing them with Granite Guardian for production deployments needing risk detection.