Research On G.722.1 Speech Coding

Speech is an important way of efficient information exchange between people.In order to improve the efficiency of speech transmission or save storage space,it is usually necessary to compress the speech signal.Speech coding techniques have been widely used in communication networks,consumer electronics,digital entertainment,national defense and military fields.The G.722.1 speech codec is a wideband speech coding standard with low complexity introduced by the International Telecommunication Union.This codec standard mainly uses transform domain coding technique,which can encode the speech with300-4000 Hz and the music within 7kHz.MELP speech codec is a low bite rate speech compress standard with a rate of 2.4kbps.Although the G.722.1 and MELP speech codecs have been used in practice,their performances are obviously degraded in the case of network packet loss.In order to improve the speech quality of codec,this thesis studies the G.722.1 and MELP speech coders.The main work is as follows:(1)A multiple description coding method based on the G.722.1 encoder is proposed.This method applies the multiple description coding idea to construct a complementary encoder of the G.722.1 encoder.Then,at the coding end,a frame of the speech is encoded by the G.722.1 encoder and its complementary encoder,respectively,while at the decoding end,when any one of the speech streams is received,it is decoded by the G.722.1 decoder;when the two speech streams are received,them are decoded jointly by the G.722.1 decoder and its complementary encoder.Thus,the speech quality is improved obviously.The simulation results show that this method has good anti-packet loss ability and obtains high speech quality.(2)In order to improve the quality of the decoded speech,a post-processing method of the G.722.1 encoder based on LSTM network is proposed.This method uses the long short-term memory(LSTM)network to learn the relationship between the original and coded speech cepstrum parameters of the G.722.1 encoder;Then the decoded speech is passed through the trained LSTM to enhance its speech quality.Finally,the original and enhanced decoded speeches are added in the frequency domain.The experimental results show that this method fills the gap in the 7kHz-8kHz frequency band of the original decoded speech and improves the quality of the decoded speech.(3)For MELP encoder,the influence of quantization error of encoding parameters,such as line spectrum frequency,pitch period,and residual harmonic amplitude,on the quality of the decoded speech is analyzed,and the corresponding experimental results are given.