| Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 93 === In this thesis, we study the problem of improving the speech fidelity in Internet telephony under the restriction that only a small data rate overhead is tolerable. Our way is to add high frequency components into the speech signals. Based on the speech production model of linear prediction, in the sender side, we analyze the voicing strength, gain and envelope of highband part of speech for every 10ms frame. Then, these analyzed parameter values are quantized and packed into 16 bits. This data rate is one fifth of the original rate of ITU G729 codec. In the receiver side harmonic and white noise generators are used to generate excitation signals, respectively. These two are mixed according to the voicing strength parameter. Then, the mixed excitation signal, after gain adjusting, is filtered by the all pole model constructed in terms of linear prediction coefficients. The obtained highband signal part is added with the lowband signal part generated by G729 decoder to synthesize the desired wideband speech.
To evaluate the quality of the synthesized wideband speech, we first port the source code of G729 codec into the program of winRTP which is used for Internet telephony. Then, we implement the needed program modules for processing the highband part of speech. After integration of the system, we conduct a subjective perception tests in which the quality of the speech of original G729 and our wideband speech are measured. The result of the tests show that the quality of our bandwidth augmented speech has apparent improvement over the original G729 speech.
|
|---|
