Apple has once again made a breakthrough in its research, improving speech recognition capabilities and reducing resource consumption while retrieving information from large databases. One of the challenges in using ASR systems is identifying rare and user-specific terms, which Apple has addressed by using NCB to retrieve information from external databases. However, this requires significant computational and memory resources. To solve this problem, Apple has introduced a two-stage approach using vector quantisation to select relevant biasing entries and a ‘cross attention’ mechanism to apply them for improved speech recognition. This has resulted in a 20% reduction in computational time and a 71% error reduction rate. The technique also processes millions of entries without compromising recognition quality. This research indicates Apple’s commitment to enhancing AI and keeping it on-device, potentially paving the way for integration of Apple Intelligence into older iPhones. With voice assistants still far from perfect, this development is a promising step towards improving their performance.
