A Novel Incremental Dictionary Learning Method for Low Bit Rate Speech Streaming

Speech streaming, which is widely used nowadays, cost a huge amount of transfer bandwidth and storage space. It is significant to compress them with as few bits as possible, meanwhile keep the voice clear and meaning unchanged. According to speech contexts, the proposed method can dynamically adapt to speech stream of any speaker by appending atoms to dictionary. Furthermore, in order to smoothly represent the amplitude envelopes shifting over the frequency, the dictionary is extended by Hilbert transform. The upper bounds of weights of atoms are constrained, so they can be quantized in practical applications. Experimental results show the advantages of our method. When the minimum reconstruction accuracy is 99.8%, which is applicable to general voice communications, the space saving is over 99%. Our method can adapt to the application with extreme bandwidth/storage limitation and large scale dataset, meanwhile keep reasonable perception quality.