IBM adds multilingual 2B model to Granite Speech, updates two checkpoints

IBM released a new Granite Speech 5.0 2B model and revised its 470M-parameter TurboCTC and Granite Speech 4.1 2B Plus checkpoints. Together, the three releases expand IBM’s open speech stack across multilingual transcription, translation and speaker-attributed audio.

All three models carry Apache 2.0 licenses. The new 2B model handles speech recognition and bidirectional translation involving English, French, German, Spanish, Portuguese and Japanese; its model card also lists English-to-Italian and English-to-Mandarin translation. IBM says the model works with Transformers, vLLM, llama.cpp and mlx-audio.

IBM said it trained the new model on 390,000 hours of public audio, supplemented by synthetic data for Japanese recognition, keyword-biased recognition and speech translation. The model supports language-ID prompts and keyword lists of up to 2,000 terms. IBM reported a 5.06% average English word error rate on public short-form test sets and ranked the model second among the open-weight systems it compared. The company labeled those October 15 results unofficial.

The revised Granite Speech 5.0 470M TurboCTC expands an earlier English-only checkpoint to seven languages and produces punctuated, capitalized transcripts. The encoder-only model decodes audio without generating transcript tokens one at a time, a design intended for low-latency transcription on laptops, phones and other edge devices. IBM attributed a roughly 1.6-fold throughput increase—to 19,870 times real time from about 12,600—to a compilation change that reuses one model block across the stack while maintaining the same accuracy level.

The revised Granite Speech 4.1 2B Plus retains its architecture, prompt format and API, and IBM describes it as a drop-in replacement for the previous checkpoint. It provides prompt-controlled speaker attribution and word-level timestamps in five languages, but does not add punctuation or capitalization. IBM said the update cut long-form word error rates by 24%, reduced errors on 20-term keyword lists by 45% and more than halved average timestamp shift.