Granite-Speech-5.0-470M-TurboCTC

IBM released Granite-Speech-5.0-470M-TurboCTC on October 21, 2026 as a multilingual update to the English-only model first published on August 25, 2026. The model is an encoder-only conformer acoustic encoder with block self-attention, self-conditioning and temporal downsampling, feeding an output layer of 16,384 multilingual BPE units. It was trained on roughly 153,000 hours of multilingual audio from public corpora using Connectionist Temporal Classification, and inference is non-autoregressive with greedy decoding. Supported languages are English, French, German, Spanish, Portuguese, Italian and Japanese, and the model emits punctuated and capitalized transcripts. IBM positions it for enterprise low-latency, high-throughput transcription, and notes it drops capabilities of the earlier LM-equipped Granite Speech models such as speech translation and keyword biasing. It is licensed under Apache 2.0.