Better on-device dictation and cleanup for Mac with Model 4 (www.epilude.com)

Epilude has announced the release of Model 4, an upgraded on-device dictation system for Mac that enhances the Local Mode service while prioritizing user privacy. This fourth-generation model maintains the same 1.5 GB size and processing speed as its predecessor but boasts a significant quality improvement, reducing critical meaning errors by 56%. The system operates entirely offline, meaning users' audio and text are never sent to external servers. Model 4's underlying architecture features a revised cleanup model, built on the Qwen family, that effectively refines speech-to-text outputs, improving handling of self-corrections, dates, mixed-language speech, and the casual tone of informal conversation. The advancements in Model 4 are particularly significant for the AI/ML community, as they focus on high-quality performance without compromising user data privacy. With 78 out of 90 internal benchmark scenarios passed, it showcases a robust method for grading dictation models, emphasizing zero-tolerance for critical errors that alter meaning. Notably, the new model’s weight-averaging approach in training has proven successful in mitigating failure modes, positioning Model 4 as a reliable choice for users valuing both quality and privacy in dictation technology. Future developments will aim to enhance handling of longer inputs and expand multilingual support, reflecting Epilude’s commitment to continuous improvement.