How Speech Recognizers Use Expected Answers to Speed Up

This patent describes how a speech recognition system can get faster and more accurate by first checking if what you said matches what it expects you to say, instead of listening for everything.

What it covers

This patent describes a method for a speech recognition system to process user input more efficiently. During a conversation, or "speech dialog," the system identifies a set of highly probable user responses, called the "expected response," which is a smaller part of all possible answers (the "grammar of possible responses"). Instead of comparing the sounds it hears (the "acoustic features") to every single possible word, the system first compares these sounds *only* to the sound pattern (the "model") of the expected response (Claim 1). If this initial comparison yields a high enough "confidence factor" compared to an "acceptance threshold," the system quickly accepts the expected response. If not, it can then proceed to compare the input to models for other possible responses (Claim 3). For example, if a voice assistant asks "Yes or No?" it would first check if the user's speech matches the models for "yes" or "no" before searching for other words.

What it doesn't cover

• —Speech recognition systems that always compare input speech to all possible responses in a grammar simultaneously without prioritizing expected responses.
• —Systems that do not have a pre-determined "expected response" known before the user speaks.
• —Speech recognition that doesn't use a "confidence factor" and "acceptance threshold" for the expected response.
• —Systems that do not perform an initial, direct comparison *only* to the expected response model before considering other options.

The clever bit

The core innovation is the initial, focused comparison of incoming speech *only* against the sound model of the *expected* response. This avoids the computational cost of a full search across all possible responses, making the system faster and more efficient when the user's input aligns with expectations.

Why it matters

This approach significantly improves the efficiency and accuracy of speech recognition, particularly in structured conversational flows like those found in voice assistants or industrial voice-picking systems. By focusing the initial search on the most likely answers, it reduces the computational load and the chance of misinterpretations. This method is valuable in environments where quick and reliable voice interactions are critical, making voice-controlled systems more practical.

Real-world examples

• 1.Voice-controlled inventory management systems (e.g., in warehouses)
• 2.Interactive Voice Response (IVR) systems for customer service
• 3.Smart home voice commands where the system expects a specific action (e.g., "on" or "off")
• 4.In-car voice commands for navigation or media control