Extremely Fast and Accurate Transcription with Granite Speech 5.0 Turbo CTC
huggingface ·
IBM has released two compact English speech recognition models in the Granite Speech family, each with 470 million parameters. The models, granite-speech-5.0-470m-turboctc and its non-commercial variant, achieve extremely high throughput of over 12,600 RTFx on an Nvidia H200 GPU, allowing them to transcribe more than 3.5 hours of speech per second. They use an encoder-only architecture with 16 Conformer blocks, chunkwise attention, and connectionist temporal classification (CTC) loss, operating at a lower token rate than previous models. The non-commercial model is trained on additional data and carries a CC-BY-NC-SA-4.0 license, while the other is Apache 2.0 licensed. On the OpenASR leaderboard public test sets, the non-commercial model scored an aggregate 4.85% word error rate, slightly better than the 5.00% of the Apache 2.0 model. Both models also ranked among the fastest and most accurate on the far-field speech recognition leaderboard. They are designed for edge-device speech-to-text tasks and are natively supported in the transformers library.
huggingface