| Keyword spotting is an important interface of human-computer speech interaction,which plays an important role in human life.With the development of various portable devices and the need for more diversified scenes in daily life,it has become a research hotspot to study how to identify voice keywords that can be flexibly changed in the context of the difficulty in collecting labeled training data and the limited computing power of portable devices.At present,the main solution is to use few-shot speech keyword spotting,but there are still some problems to be solved.Firstly,speech keywords are a kind of speech signal with low discrimination.Some different speech keywords may show similar pronunciation and the same speech keywords may show different pronunciation.These samples are easy to be over considered or ignored when they are mapped into the metric space in the few-shot learning method based on metric learning,resulting in overfitting of these samples,which affects the recognition accuracy;In addition,when the speech keywords are mixed with noise,the interference of noise is more prominent,which affects the performance of speech keyword spotting with fewer samples,because there are fewer trainable labeled samples sent into the model each time,however the traditional noise processing method requires a lot of computing power,which is not suitable for keyword spotting in a few-shot scenarios.In view of the above difficulties and challenges,this thesis proposes a few-shot speech keyword spotting method based on multi-level distance regularization and signal-to-noise(SNR)distance regularization by combining the method based on distance measurement constraints and measuring learning,improving few-shot keyword spotting performance from the perspective of difficult examples and noise environment.The main contents and innovations of this thesis are as follows:(1)Few-shot keyword spotting method based on multi-level distance regularization.Aiming at the problem of overfitting caused by the lack of training samples in the few-shot learning scenario,which leads to the over consideration or neglect of some special speech keyword samples(which are very similar to the samples of other classes or very different from the samples of their own classes),it is proposed to combine the few-shot learning method based on metric learning with the multi-level distance regularization(MDR)method,and the latter is used as an additional loss function to jointly train the few-shot keyword spotting model.Due to the hierarchical constraint of MDR on the metric space,the metric space after training has the ability to distinguish more similarities at different levels,and thus has better generalization and distinguishability.Experiments on the Google speech commands dataset show that MDR helps the model to obtain a metric space with better generalization ability.The accuracy of the proposed method has been improved under the combination of different few-shot tasks and embedding networks,with the improvement range of 0.5% ~ 31.8%,which is higher than the best few-shot keyword spotting method at present.(2)Few-shot keyword spotting method based on SNR distance regularization.Aiming at the problem that the existing methods need to consume a lot of resources to perform noise reduction processing for few-shot keyword spotting in noisy environments,which is not suitable for the needs of few-shot scenarios,a SNR distance regularization method is proposed to improve the performance of few-shot speech keyword spottin in noisy environments.The input speech samples to be trained are mixed with noise data of different SNR ratios for sample expansion,and interval constraints are made according to different SNR,finally forming a metric space with better robustness to noisy speech keywords.Experiments on the Google Speech Commands dataset show that the proposed method for few-shot speech keyword spotting based on SNR distance regularization performs well in noisy environments,compared to the few-shot speech keyword spotting method without any processing.The average improvement is about 8.68%,which effectively proves that the proposed method has a certain effect.(3)Design and implement a prototype system for few-shot keyword spotting.Using Python and Matlab mixed programming,designs and implements a keyword spotting prototype system in a few-shot environment.The system includes speech data preprocessing,loading few-shot keyword spotting model and recognizing keywords,and visual analysis modules.The effectiveness of the proposed method is demonstrated through a prototype system. |
