What IBM announced
IBM’s October 21 Granite Speech announcement introduces granite-speech-5.0-2b and describes revisions to two existing models. These tools turn recordings into text; their roles include transcription, translation, and identifying speaker turns.
The 2B model targets multilingual transcription and speech translation. The smaller granite-speech-5.0-470m-turboctc focuses on fast transcription, adding languages and native punctuation. Granite-speech-4.1-2b-plus targets multi-speaker recordings and word-level timing. Its card says it does not provide punctuation and capitalization, so check the specific output you need.
The model-card dates differ: October 20 for the new 2B model, October 21 for the multilingual TurboCTC revision, and October 19 for the Plus update. Keep those artifact dates separate from the announcement date when recording the version you evaluate.
Access and reported results
The public repositories list model weights and Apache2.0 licenses. IBM reports improved accuracy, speed, and speaker or timing results in its comparisons. Those are provider evaluations. We have not downloaded the weights, run the models, or independently measured performance. Available code and small parameter counts still require compatible software, memory, and hardware.
Our take
For creators and experienced technicians, a transcript is a useful draft to check against the recording. Try a short clip you have permission to use, including names, a price, and an unclear word. Compare the text with the audio before using it for captions, notes, or translation.
For meetings, check who said each important sentence and whether the timing fits. Test quiet and noisy clips separately. Keep uncertain words visibly uncertain; a polished transcript can still change a number or assign a statement to the wrong person.