IBM has announced the release of two new models in the Granite Speech family, offering high accuracy and unprecedented speed for speech recognition tasks. These compact models, with 470M parameters, achieve over 12,600 RTFx on an NVIDIA H200 GPU, allowing for the transcription of more than 3.5 hours of speech in one second using batched inference. The models are designed for efficient performance on edge devices and are available under different licensing terms.
The two models, granite-speech-5.0-470m-turboctc and granite-speech-5.0-470m-turboctc-nc, differ primarily in their training data and licensing. The non-commercial model, granite-speech-5.0-470m-turboctc-nc, is trained on additional data and carries a CC-BY-NC-SA-4.0 license, while the Apache 2.0 licensed model, granite-speech-5.0-470m-turboctc, is trained on a smaller dataset. Both models perform well on far-field speech recognition, with the non-commercial model ranking fifth and the Apache 2.0 model ranking ninth on the FFASR Leaderboard as of 25 August 2026.
The Granite 5.0 Speech models are encoder-only models, which provide a smaller memory footprint and faster throughput compared to prior models. They use a stack of 16 Conformer blocks, self-conditioning at the output of the 8th block, and chunkwise attention to avoid quadratic scaling with sequence length. These models are optimized for speech-to-text tasks and are supported natively in the transformers library.
Source: huggingface
