Research On Speech Verification For Mobile Devices

After more than half a century of development, speech recognition technology has become increasingly mature, especially speech input methods, voice command and conversion techniques between speech and text. Then with the popularity of smartphones in recent years, speech recognition technology has entered into the era of mobile devices. However, due to the limitations of speech verification on the mobile terminal, such as short utterance, less training set, complicated environmental noises and diversity of mobile terminals, speech verification on the mobile devices is not relatively accurate, hindering the marketing and application of speech verification in the field of mobile devices. In addition, majority of recent studies always focused on speech contents recognition and speech synthesis, while there has not been a big breakthrough on the applications of speech verification based on mobile devices. In this dissertation, three algorithms or architectures are proposed from three aspects, namely, speech pre-processing, speech verification architecture and speech contents, to solve the low-accuracy problem on the speech verification on the mobile devices, and then validate the analysis.In this dissertation, from speech pre-processing aspect, effective components extraction algorithm for speech based on short-time energy is proposed in the speech pre-processing, which is modified by the conventional endpoint detection algorithm and which can remove the mute noise section to extract the effective components in the speech signal. Experiments show that this algorithm can efficiently reduce the amount of speech signal and weaken the noise interference. Then, this dissertation presents a multistage verification architecture, combining feature extraction methods based on wavelet algorithm and DTW matching algorithm, to achieve a better verification. Feature extraction algorithm based on wavelet transform can efficiently extract multi-degree feature parameters of speech signal in time domain and frequency domain meanwhile and can also obtain more characteristic parameters from small training set. The experimental results show that the architecture can achieve 3% or more of the ascension for different mobile devices. Finally, this dissertation proposes a modified speech verification system based on Arabic numeral strings. The experimental results show that for 10 digits, the number 3,4,5,6 and 8 five digits have better recognition of a speech. According to the results of the experiment, this project make an optimization and impro vement for the speech verification system based on 8 numeric strings, which obtains at least 5% improvement.This topic put forward three contents, from different angles of speech verification system to improve the overall recognition performance. The syst em implementation also proves the feasibility of the system at last. The whole framework based on the mobile speech verification system has higher reference value for the subsequent research.