Experts at Apple are exploring the potential use of artificial intelligence to identify when a person is communicating with a device like an iPhone, aiming to eliminate the necessity for a specific wake word such as “Siri,” as per a report released recently.
The study, available on Arxiv and not yet peer-reviewed, involved training a sophisticated vocabulary model using speech data from smartphones and ambient noise to detect patterns indicating user interactions with the device. Leveraging a modified version of OpenAI’s GPT-2, the researchers, six of whom are affiliated with Apple, including three from the Siri team, emphasized the model’s ability to operate effectively on devices like smartphones due to its lightweight nature. The research utilized over 129 hours of data for training, although the origin of the recordings remains unspecified.
Reportedly, the model exhibited promising results, surpassing the accuracy of audio-only or text-only models, with performance improving as the model size increased. While the study raises questions about the future of the “Hey Siri” wake phrase, Apple’s intentions regarding its elimination remain unclear.
Apple and the researchers mentioned in the paper have yet to respond to requests for comments on the findings. The current Siri functionality involves storing minimal audio until triggered by the wake phrase, a design that has raised concerns about privacy implications and constant audio monitoring.
Privacy advocates have previously criticized Apple’s handling of audio data, particularly after reports in 2019 revealed contractors listening to private conversations via Siri recordings. In response, Apple implemented policy changes to enhance user control over data usage for Siri improvements. Despite the potential convenience of removing the wake phrase, experts like Jen King caution that it could reduce transparency and awareness of device interactions.
The discussion surrounding the research paper signals Apple’s strategic shift towards integrating more AI capabilities into its products to compete with tech giants like Amazon, Google, and Facebook. Reports suggest Apple’s development of a generative AI model named MM1 for text and image applications, akin to OpenAI’s ChatGPT, as well as potential collaborations with Google and Baidu for AI integration in iPhones.
