Sectioned Memory Networks for Online Keyword Spotting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies are inefficient in detecting words in continuous speech, relying on phoneme detection and hidden Markov models, which lack accuracy and speed.
Innovation Solution
The use of sectioned memory networks and recurrent neural networks to directly identify keywords in speech by generating blocked feature vector sequences and processing them to determine the presence of specific words without requiring phonetic composition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If hidden Markov models and neural networks are used to compute features of speech, then speech recognition can be performed, but accuracy and speed of detecting words in speech deteriorates
Solution Approach 1:
The speech signal is divided into overlapping frames of fixed duration, with each frame processed independently by the neural network to detect keyword presence. This segmentation allows parallel processing of multiple time segments, improving both accuracy through focused analysis and speed through concurrent computation.
Solution Approach 2:
The patent replaces the traditional hidden Markov model approach with a neural network-based system that directly processes spectral features. This substitution eliminates the need for phoneme-level analysis and secondary searches, providing more accurate and faster keyword detection through direct pattern recognition in the spectral domain.
2Reliability
If phoneme detection and composition are used to identify words, then speech recognition can be performed, but processing time increases
Solution Approach 1:
The patent extracts and processes only the relevant spectral features directly from speech frames, bypassing the intermediate phoneme detection step. By taking out the essential keyword detection function and processing it directly on spectral data, the system eliminates unnecessary processing stages and reduces overall processing time while maintaining detection capability.
Solution Approach 2:
The neural network is pre-trained to recognize keyword patterns directly from spectral features, performing the detection action in advance within each processed frame. This preliminary action within each frame eliminates the need for subsequent composition and search operations, significantly reducing processing time.
Data Source
AI summary
Systems, methods, and computer program products to detect a keyword in speech, by generating, from a sequence of spectral feature vectors generated from the speech, a plurality of blocked feature vector sequences, and analyzing, by a neural network, each of the plurality of blocked feature vector sequences to detect the presence of the keyword in the speech.