A new system described on arXiv, named oollim, is designed to evaluate L2 English pronunciation on a consumer computer. The abstract states that it is locally executable, meaning it can run on a typical gaming PC without relying on cloud services or high-end infrastructure.
The system assesses three aspects of speech: accuracy, fluency, and prosody. It does so by combining pretrained representations from XLS-R, XEUS, and PhoneticXeus with reference-text information, presumably to compare the learner's spoken output against expected content.
The source is brief, and the abstract is truncated, so details on training data, evaluation benchmarks, or performance metrics are not provided. The main takeaway is the feasibility of running such a multi-model speech evaluation pipeline entirely on local hardware.