IBM Launches High-Speed Granite Speech 5.0 Turbo CTC Models

IBM released two new compact, 470-million-parameter English speech recognition models in the Granite Speech family. [1] These models achieve speeds of over 12,600 RTFx on an NVIDIA H200 GPU. [1] Using batched inference, the systems can transcribe more than 3.5 hours of speech in one second. [1] A streaming speech recognition WebGPU demo is available for Chrome and Edge browsers. [1]

The granite-speech-5.0-470m-turboctc model uses a smaller training dataset and carries an Apache 2.0 license. [1] This model scores a 5.00% aggregate Word Error Rate. [1] The granite-speech-5.0-470m-turboctc-nc version includes additional training data and is released under a CC-BY-NC-SA-4.0 license. [1] The noncommercial model achieves a 4.85% aggregate Word Error Rate, showing an advantage on the SPGI Speech test but a disadvantage on the chunked Earnings22 test. [1]

The focus on speed combined with low parameter count positions the Granite Speech 5.0 Turbo CTC models as highly efficient transcription tools. By offering both Apache 2.0 and noncommercial variants, IBM provides options tailored to enterprise integration versus experimental research. The models compete effectively in terms of Word Error Rate while prioritizing sheer throughput on hardware like the NVIDIA H200, matching leading open-source alternatives. What the sources don't address: How the models handle complex acoustic environments or languages beyond English.