The FS-1016 CELP standard.
a. Obtain the MATLAB software for the FS-1016 CELP from the Book website. Use the following wave files: Ch4Sp8.wav and Ch4Au8.wav.
Deliverable 1:
Give the plots of the entire input and the FS-1016 synthesized output for the two wave files. Comment on the quality of the two synthesized wave files. In particular, give more emphasis on the Ch4Au8.wav and give specific reasons why the FS-1016 does not synthesize Ch4Au8.wav with high quality. Listen to the output files and provide a subjective evaluation. The CELP FS1016 coder scored 3.2 out of 5 in government tests on a MOS scale. How would you rate its performance in terms of MOS for Ch4Sp8.wav and Ch4Au8.wav? Also give segmental SNR values for a voiced/unvoiced/mixed frame and overall SNR for the entire record. Present your results as Table 4.4 with appropriate caption.
b. Spectrum analysis: File CELPANAL.M, from lines 134 to 159 . Valuable comments describing the variable names, globals, and inputs/ outputs are provided at the beginning of the MATLAB file CELPANAL.M to further assist you with understanding the MATLAB$$
\text { Table 4.4. FS-1016 CELP subjective and objective evaluation. }
$$FIGURE CANT COPY
program. Specifically, some of the useful variables in the MATLAB code include snew-input speech buffer, fcn-LP filter coefficients of $1 / A(z)$, rcn-reflection coefficients, newfreq-LSP frequencies, unqfrequnquantized LSPs, newfreq-quantized LSPs, and 1sp-interpolated LSPs for each subframe.
Deliverable 2:
Choose Ch4Sp8.wav and use the voiced/unvoiced/mixed frames selected in the previous step. Indicate the frame numbers. FS-1016 employs 30-ms speech frames, so the frame size is fixed $=240$ samples. Give time-domain (variable in the code 'snew') and frequency-domain plots (use FFT size 512; include commands as necessary in the program to obtain the FFT) of the selected voiced/unvoiced/mixed frame; Also plot the LPC spectrum using
figure, freqz $(1, f c n)$
Study the interlacing property of the LSPs on the unit circle. Note that in the FS-1016 standard, the LSPs are encoded using scalar quantization. Plot the LPC spectra obtained from the unquantized LSPs and quantized LSPs. You have to convert LSPs to LPCs in both the cases (unquantized and quantized) and use the freqz command to plot the LPC spectra. Give a $z=$ domain plot (of a voiced and an unvoiced frame) containing the pole locations (show as crosses ' $x$ ') of the LPC spectra, and the roots of the symmetric (show as black circles 'o') and asymmetric (show as red circles 'o') LSP polynomials. Note the interlacing nature of black and red circles, they always lie on the unit circle. Also note that if a pole is close to the unit circle, the corresponding LSPs will be close to each other. In the file CELPANAL.M; line 124, high-pass filtering is performed to eliminate the undesired low frequencies. Experiment with and without a high-pass filter to note the presence of humming and low-frequency noise in the synthesized speech.
c. Pitch analysis: File: CELPANAL.M; from lines 162 to 191 .
Deliverable 3:
What are the key advantages of employing subframes in speech coding (e.g., interpolation, pitch prediction?). Explain, in general, the differences between long-term prediction and short-term prediction. Give the necessary transfer functions. In particular, describe what aspects of speech each of the two predictors captures.
Deliverable 4:
Insert in file CELPANAL.M; after line 182
$$
\text { tauptr }=75 \text {; }
$$
Perform an evaluation of the perceptual quality of synthesis speech and give your remarks. How does the speech quality change by forcing a pitch to a predetermined value? (Choose different tauptr values, 40, 75, and 110.)