Siri's Lag Fixed?! 🤯 Faster Responses Revealed! ✨
Tech
🎧
Apple researchers are investigating improvements to text-to-speech systems, focusing on reducing delays in responses. A recently published paper details a proposed change to current systems, which heavily rely on autoregression—generating speech token by token. The team, in collaboration with Tel Aviv University researchers, suggests replacing this approach with “Acoustic Similarity Groups,” or ASGs, containing perceptually similar sounds. This shift aims to mitigate delays by allowing the system to consider the overall sound rather than isolating individual tokens. The research suggests that even small delays can disrupt the natural flow of voice interactions. Ultimately, the goal is to refine spoken responses, tailoring them to user preferences and environmental context.
FASTER SPEECH GENERATION
The Apple Intelligence team, in collaboration with Tel Aviv University, is pioneering a method to dramatically reduce the delay between a user’s request and Siri’s spoken response. This research highlights that even minor delays within AI voice interfaces can significantly disrupt the flow of conversation and diminish the perception of responsiveness. Current text-to-speech systems typically generate speech by processing text as a sequence of tokens, representing extremely short sound snippets measured in milliseconds. These tokens correspond to phonetic units, and slight mismatches can lead to odd pronunciations, misplaced emphasis, or occasional mispronunciations – issues frequently encountered with Siri.
ACOUSTIC SIMILARITY GROUPS
A key challenge lies in conversational settings where pauses exceeding a fraction of a second can make an assistant feel slow and disengaged. Existing systems heavily rely on autoregression, generating speech tokens sequentially and narrowing down choices based on previously selected tokens. This approach exacerbates the problem of ignoring acoustic similarity between sounds and increases the risk of “erroneous acceptances,” where a technically correct token is chosen but sounds unnatural to human listeners. The sequential nature of autoregression also limits the ability to skip ahead or parallelize parts of the process, directly impacting speech generation speed.
REDEFINING TOKEN MATCHING
Apple’s proposed solution involves replacing strict, exact token matching with a broader, probabilistic approach. This entails grouping tokens into what they term Acoustic Similarity Groups (ASGs). ASGs comprise “perceptually similar sounds,” recognizing that humans perceive closely related sounds even if they aren't identical at a technical level. By evaluating groups of tokens simultaneously, the system avoids the pitfalls of autoregression and drastically improves the speed and naturalness of speech generation.
PERSONALIZED VOICE INTERACTION
Furthermore, Apple researchers are investigating ways to tailor spoken responses to individual user preferences and environmental factors. This includes adjusting tone, pacing, and clarity based on context. Combined with faster speech generation and the ASG approach, the ultimate goal is to create voice assistants that feel less mechanical and more responsive, offering a gradual shift toward conversations that are smoother, quicker, and more closely aligned with human speech. Apple has introduced anew iPhone privacy settingthat blurs location data shared with wireless carriers.
This article is AI-synthesized from public sources and may not reflect original reporting.