Keyword Detection Using Feature Vector Scoring and Posterior Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice recognition systems for mobile devices face challenges in efficiently recognizing keywords with low latency and power consumption while maintaining flexibility and adaptability to different speakers.
Innovation Solution
The system employs a data processing apparatus that generates output feature vectors for audio frame vectors, using a neural network or Gaussian mixture model to determine the presence of keywords by combining scores from acoustic modeling, and a posterior handling module to make binary decisions on keyword detection, minimizing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network or Gaussian mixture model is used for keyword recognition, then measurement precision is improved, but use of energy increases
Solution Approach 1:
The system segments the audio processing task into distinct stages: audio frame extraction, feature vector generation, scoring against expected event vectors, and posterior handling. This segmentation allows the device to process audio in manageable chunks rather than continuously analyzing entire audio streams, reducing overall power consumption while maintaining recognition accuracy through focused processing at each stage.
Solution Approach 2:
The system performs partial action by generating scores for only the most relevant expected event vectors rather than exhaustively evaluating all possible keywords. The posterior handling module then makes decisions based on these partial scores, avoiding the need to process every possible keyword candidate with full neural network computation, thus reducing energy usage while maintaining sufficient recognition precision.
2Reliability
If audio processing is performed continuously to improve keyword detection reliability, then productivity is improved, but use of energy increases
Solution Approach 1:
The system implements periodic action by processing audio in discrete frames at specific intervals rather than continuously analyzing the audio stream. Audio is divided into frames, each processed independently to generate feature vectors and scores. This periodic processing approach maintains reliable keyword detection by consistently sampling the audio stream while significantly reducing power consumption compared to continuous processing.
Solution Approach 2:
The system performs preliminary action by pre-computing and storing expected event vectors for known keywords before runtime. During audio processing, the system compares extracted audio features against these pre-computed vectors rather than performing full keyword recognition from scratch. This preliminary preparation enables faster, more reliable detection while reducing the computational energy required during active audio processing.
3Measurement precision
If the system processes all audio frames to ensure accurate keyword detection, then measurement precision is improved, but loss of time increases
Solution Approach 1:
The system extracts only the most relevant features and scores from audio frames for keyword detection, rather than processing and analyzing all audio frame data in full detail. The feature extraction module identifies and extracts discriminative characteristics, and the scoring module focuses computation on generating relevance scores for expected event vectors. This extraction approach maintains detection accuracy by preserving key information while reducing processing time and latency.
Solution Approach 2:
The system implements skipping by rapidly processing audio frames through the feature extraction and scoring pipeline, moving quickly from raw audio input to keyword detection decisions. The streamlined architecture allows the system to skip unnecessary computational steps and directly generate detection results from extracted features, reducing overall processing time and latency while maintaining sufficient accuracy for wake-word detection applications.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for key phrase detection. One of the methods includes receiving a plurality of audio frame vectors that each model an audio waveform during a different period of time, generating an output feature vector for each of the audio frame vectors, wherein each output feature vector includes a set of scores that characterize an acoustic match between the corresponding audio frame vector and a set of expected event vectors, each of the expected event vectors corresponding to one of the scores and defining acoustic properties of at least a portion of a keyword, and providing each of the output feature vectors to a posterior handling module.