IBM Releases Granite Speech 5.0 TurboCTC, a 470M Apache-2.0 ASR Model Claiming 12,600 RTFx on One H200
IBM published two 470M-parameter English speech recognition models on August 25, granite-speech-5.0-470m-turboctc under Apache 2.0 and a noncommercial counterpart, both using 16 conformer blocks trained with CTC on a 16,384 BPE head, temporal subsampling by 8, 128-frame block attention and self-conditioned CTC from the middle layer. IBM reports aggregate throughput above 12,600 RTFx on a single NVIDIA H200, roughly 3.5 hours of audio per second, with 5.00% aggregate WER for the open variant and 4.85% for the noncommercial one on OpenASR English short-form sets. Those accuracy numbers are vendor-reported and IBM labels them unofficial pending the leaderboard update, though inference ran on Hugging Face jobs infrastructure with leaderboard tooling.
↳ Follow the thread