Real-Time Keyword Spotting via Posterior Probability and Garbage Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current keyword spotting systems in automatic speech recognition face challenges such as high complexity, error rates, and computational costs, particularly in real-time applications, due to reliance on phonetic decoding, garbage models, and large vocabulary continuous speech recognition, which hinder accurate and efficient keyword detection across multiple languages.
Innovation Solution
A system that computes posterior probabilities of audio streams matching keyword models, using a recognition engine to compare audio data with keyword models, and employs empirical metrics to reject false alarms, allowing for real-time keyword spotting with improved accuracy and reduced computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If phonetic decoding is used for keyword spotting, then keyword detection capability is improved, but system complexity and processing time increase significantly
Solution Approach 1:
The system segments the keyword spotting task into two independent stages: audio-to-text conversion using ASR, and keyword matching using garbage models. This segmentation allows each component to be optimized separately, reducing overall system complexity while maintaining detection accuracy.
Solution Approach 2:
The system uses ASR to generate a text transcription (copy) of the audio stream, which is then processed by simple garbage models for keyword detection. This copying approach avoids the complexity of direct phonetic decoding while preserving the essential information needed for accurate keyword spotting.
2Measurement precision
If phonetic decoding with language models is used, then keyword spotting accuracy improves, but computational resources and processing time increase
Solution Approach 1:
The system replaces expensive phonetic decoders with cheap garbage models that use simple phoneme sequences. These garbage models consume minimal computational resources while achieving comparable keyword detection accuracy, making them suitable for resource-constrained real-time applications.
3Adaptability or versatility
If LVCSR speech recognition engine is used for keyword spotting, then comprehensive keyword detection is achieved, but real-time processing becomes infeasible due to high computational cost
Solution Approach 1:
The system extracts only the essential keyword matching function from the complex LVCSR engine by using garbage models with phoneme sequences. This extraction retains the ability to detect keywords across multiple languages and domains while eliminating the heavy computational overhead of full speech recognition.
Data Source
AI summary
A system and method are presented for real-time speech analytics in the speech analytics field. Real time audio is fed along with a keyword model, into a recognition engine. The recognition engine computes the probability of the audio stream data matching keywords in the keyword model. The probability is compared to a threshold where the system determines if the probability is indicative of whether or not the keyword has been spotted. Empirical metrics are computed and any false alarms are identified and rejected. The keyword may be reported as found when it is deemed not to be a false alarm and passes the threshold for detection.