| Query-by-example technology is one of the important research directions in the field of keyword detection.It is widely applied in audio information retrieval,human-computer interaction,information security and so on,and has a broad application prospect.Query-by-example technology uses the template of keywords to retrieval keywords from the test speech,that is,using speech to retrieval speech.In recent years,with the rise of deep learning technology,query-by-example technology based on deep learning has also developed,which has been widely concerned by researchers.This thesis focuses on the embedding-based query-byexample system and phonetic posterior features based query-by-example keyword detection system.The system implementation and performance analysis are carried out.The main contents of this thesis are as follows:(1)This thesis studies the acoustic word embedding model,combines the multi-view method and the recurrent autoencoder network,designs the acoustic embedding model based on the multi-view method.The semi-hard negative example selection strategy is used to construct triplets,so as to optimize the network by using triplet loss.The performance of the algorithm is verified by experiments.(2)This thesis studies embedding-based query-by-example keyword detection system,analyzes the approximate nearest neighbor search based on locality sensitive hashing,and proposes an improved sliding window algorithm.At the same time,the system threshold setting method is studied,and we use multiple templates to detect keywords,and carry out the experimental test.(3)A multi-layer perceptron based phoneme recognizer is built to extract the phonetic posteriorgram.We use the frame level approximate nearest neighbor search algorithm based on locality sensitive hashing to generate the sparse similarity matrix between the template and the test speech.Then we use the two-pass repeated trajectory search combined with image processing technology for keyword detection.Finally,the system performance is improved by using multiple templates,and the performance of system is analyzed. |
