Vehicle Speech Keyword Control With Two-Step Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems for vehicle control often suffer from high error rates when recognizing multi-part keywords due to misinterpretation of keyword parts in contexts different from the intended activation, leading to unwanted system activations.
Innovation Solution
A two-step recognition process is employed, where a first part of the keyword (differentiator) is identified, followed by a second part (body) within time and data volume limits, using similarity thresholds and buffer management to reduce errors, and a continuous phonetic symbol buffer to handle pauses and noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire keyword is recognized as a whole in a single search phase, then the recognition process is simple and fast, but the error rate is high (0.1) due to misinterpretation of keyword parts in different contexts
Solution Approach 1:
The keyword recognition process is segmented into two distinct search phases: a first search phase for the differentiator part and a second search phase for the body part. This segmentation allows each part to be recognized independently with lower error rates (0.2 for differentiator, 0.1 for body), resulting in an overall error rate of 0.02, while maintaining manageable system complexity through structured phase management.
2Reliability
If the second part of the keyword is searched for indefinitely after the first part is recognized, then complete keyword recognition is achieved, but processing time increases and false activations from pauses are prevented
Solution Approach 1:
A predetermined time limit is established before the second search phase begins. This preliminary time constraint ensures that the system will terminate the search for the second part after a specific duration, preventing indefinite processing while still allowing complete keyword recognition. The time limit is set to accommodate normal speech variations while preventing false activations from pauses.
Solution Approach 2:
The search process dynamically adjusts its behavior based on real-time conditions. The system monitors the duration of the second search phase and automatically terminates when the predetermined time limit is reached, or when the phonetic sequence no longer matches the expected body part pattern. This dynamic termination prevents excessive processing time while maintaining recognition accuracy.
3Reliability
If a large buffer of phonetic symbols is maintained to capture complete keyword phrases, then recognition accuracy is improved, but memory consumption and processing overhead increase
Solution Approach 1:
The phonetic sequence processing is segmented into two phases with different buffer requirements. In the first search phase, a buffer is maintained for the differentiator part. After the differentiator is recognized, the system transitions to the second search phase where the buffer is reset and reused for the body part. This segmentation allows the system to maintain sufficient buffer capacity for accurate recognition while reducing overall memory consumption by reusing the same buffer space.
Solution Approach 2:
The phonetic symbol buffer is discarded (cleared) after the first search phase completes and the differentiator is recognized. The buffer is then recovered and reused for storing phonetic symbols during the second search phase for the body part. This discard-and-recover mechanism reduces the effective buffer size needed at any given time while maintaining the ability to store sufficient phonetic symbols for accurate keyword recognition.
Data Source
AI summary
Multi-part keywords for activating a speech operating system are recognized by receiving a phonetic sequence and checking the phonetic sequence for the presence of a first part of the keyword. Only if the first part is recognized, the phonetic sequence is checked for the presence of a second part of the keyword. The speech operating system for carrying out an action is activated only if the second part is recognized.