A Hugging Face Blog post announces two Granite Speech 5.0 automatic speech recognition models for English: ibm-granite/granite-speech-5.0-470m-turboctc and ibm-granite/granite-speech-5.0-470m-turboctc-nc. Both target speech-to-text workloads on edge devices, with benchmark results that place them among the fastest models listed in the cited comparisons.
Speed is the headline figure
Each model contains 470 million parameters. In batched inference on an NVIDIA H200, their combined throughput exceeds 12,600 RTFx. In practical terms, that figure corresponds to more than 3.5 hours of speech transcribed in one second.
The measurement is tied to a specific GPU and batch-inference setup. It should therefore be read as a benchmark result, not as a universal speed claim for every edge device, workload or deployment pattern.
The models use an encoder-only architecture. According to the post, that design delivers more than 20 times the throughput of previous Granite Speech models. It keeps the systems focused on converting speech into text rather than extending them into a broader speech-language workflow.
The licensing choice changes the deployment equation
The ibm-granite/granite-speech-5.0-470m-turboctc-nc model was trained with additional data and records a 4.85% aggregate word error rate on the public short-form English test sets used by the OpenASR leaderboard. The Apache 2.0 version, trained on a smaller dataset, records 5.00% on the same tests.
The reported accuracy gap is modest, but the licensing difference has a direct effect on deployment decisions. The additional-data model is distributed under CC-BY-NC-SA-4.0, while the smaller-data model uses Apache 2.0. Teams choosing between them must therefore weigh both the recognition results and the permissions attached to each model.
Leaderboard positions remain snapshots
As of August 25, 2026, the Apache 2.0 model ranked ninth for accuracy on the FFASR leaderboard, while the non-commercial model ranked fifth. The two Granite Speech 5.0 models were also the two fastest models on that leaderboard at that date.
Those positions describe the cited leaderboard at a specific moment. They do not guarantee that the rankings will remain unchanged as the leaderboard evolves, nor do they establish identical results across every device or speech scenario.
The stated scope also sets a clear boundary. These encoder-only systems are designed for English speech-to-text on edge devices, but they do not provide speech translation or keyword biasing. The announcement is consequently most relevant to teams that need high-throughput English transcription and can operate under the selected license. Projects requiring those specialised speech features would need capabilities beyond the set described here.
Granite Speech 5.0 offers two distinct deployment paths: additional training data with a non-commercial license, or a smaller training set under Apache 2.0. The public figures support a sharp throughput advantage and a measurable difference in reported recognition accuracy, but they do not remove the need to match the model and license to the intended deployment.