IBM's speech model claims 3.5 hours of audio a second

IBM released granite-speech-5.0-470m-turboctc, a 470 million parameter speech recognition model under Apache 2.0. The company reports over 12,600 RTFx on a single H200 GPU, roughly 3.5 hours of audio transcribed per second of compute.

That throughput number matters more than the model size. Most teams paying a cloud vendor per minute of audio are paying for compute they could run themselves, if IBM's 5.00% word error rate on OpenASR English short-form sets holds up outside its own testing.

The architecture stacks 16 conformer blocks trained with CTC under a 16,384 BPE token head. It subsamples the audio timeline by a factor of 8 and uses 128-frame block attention. A self-conditioned CTC pass off the middle layer feeds back into later layers. None of that is new by itself. IBM packed it into 470 million parameters and still reached that 5.00% word error rate.

IBM's own model card labels the numbers unofficial, pending an update to the public leaderboard. That's a real caveat. Vendor-reported throughput and accuracy figures often drift once independent runs replicate them on different hardware or batch sizes.