CS578- Speech Signal Processing

Size: px

Start display at page:

Download "CS578- Speech Signal Processing"

Grant Bishop
5 years ago
Views:

1 CS578- Speech Signal Processing Lecture 7: Speech Coding Yannis Stylianou University of Crete, Computer Science Dept., Multimedia Informatics Lab Univ. of Crete

2 Outline 1 Introduction 2 Statistical Models 3 Scalar Quantization Max Quantizer Companding Adaptive quantization Differential and Residual quantization 4 Vector Quantization The k-means algorithm The LBG algorithm 5 Model-based Coding Basic Linear Prediction, LPC Mixed Excitation LPC (MELP) 6 Acknowledgments

3 Digital telephone communication system

4 Categories of speech coders Waveform coders (16-64 kbps, f s = 8000Hz) Hybrid coders ( kbps, f s = 8000Hz) Vocoders ( kbps, f s = 8000Hz)

5 Categories of speech coders Waveform coders (16-64 kbps, f s = 8000Hz) Hybrid coders ( kbps, f s = 8000Hz) Vocoders ( kbps, f s = 8000Hz)

6 Categories of speech coders Waveform coders (16-64 kbps, f s = 8000Hz) Hybrid coders ( kbps, f s = 8000Hz) Vocoders ( kbps, f s = 8000Hz)

7 Speech quality Closeness of the processed speech waveform to the original speech waveform Naturalness Background artifacts Intelligibility Speaker identifiability

8 Speech quality Closeness of the processed speech waveform to the original speech waveform Naturalness Background artifacts Intelligibility Speaker identifiability

9 Speech quality Closeness of the processed speech waveform to the original speech waveform Naturalness Background artifacts Intelligibility Speaker identifiability

10 Speech quality Closeness of the processed speech waveform to the original speech waveform Naturalness Background artifacts Intelligibility Speaker identifiability

11 Speech quality Closeness of the processed speech waveform to the original speech waveform Naturalness Background artifacts Intelligibility Speaker identifiability

12 Measuring Speech Quality Subjective tests: Diagnostic Rhyme Test (DRT) Diagnostic Acceptability Measure (DAM) Mean Opinion Score (MOS) Objective tests: Segmental Signal-to-Noise Ratio (SNR) Articulation Index

13 Measuring Speech Quality Subjective tests: Diagnostic Rhyme Test (DRT) Diagnostic Acceptability Measure (DAM) Mean Opinion Score (MOS) Objective tests: Segmental Signal-to-Noise Ratio (SNR) Articulation Index

14 Measuring Speech Quality Subjective tests: Diagnostic Rhyme Test (DRT) Diagnostic Acceptability Measure (DAM) Mean Opinion Score (MOS) Objective tests: Segmental Signal-to-Noise Ratio (SNR) Articulation Index

15 Measuring Speech Quality Subjective tests: Diagnostic Rhyme Test (DRT) Diagnostic Acceptability Measure (DAM) Mean Opinion Score (MOS) Objective tests: Segmental Signal-to-Noise Ratio (SNR) Articulation Index

16 Measuring Speech Quality Subjective tests: Diagnostic Rhyme Test (DRT) Diagnostic Acceptability Measure (DAM) Mean Opinion Score (MOS) Objective tests: Segmental Signal-to-Noise Ratio (SNR) Articulation Index

17 Quantization Statistical models of speech (preliminary) Scalar quantization (i.e., waveform coding) Vector quantization (i.e., subband and sinusoidal coding)

18 Quantization Statistical models of speech (preliminary) Scalar quantization (i.e., waveform coding) Vector quantization (i.e., subband and sinusoidal coding)

19 Quantization Statistical models of speech (preliminary) Scalar quantization (i.e., waveform coding) Vector quantization (i.e., subband and sinusoidal coding)

20 Linear Prediction Coding, LPC Classic LPC Mixed Excitation Linear Prediction, MELP Multipulse LPC Code Excited Linear Prediction (CELP)

21 Linear Prediction Coding, LPC Classic LPC Mixed Excitation Linear Prediction, MELP Multipulse LPC Code Excited Linear Prediction (CELP)

22 Linear Prediction Coding, LPC Classic LPC Mixed Excitation Linear Prediction, MELP Multipulse LPC Code Excited Linear Prediction (CELP)

23 Linear Prediction Coding, LPC Classic LPC Mixed Excitation Linear Prediction, MELP Multipulse LPC Code Excited Linear Prediction (CELP)

24 Outline 1 Introduction 2 Statistical Models 3 Scalar Quantization Max Quantizer Companding Adaptive quantization Differential and Residual quantization 4 Vector Quantization The k-means algorithm The LBG algorithm 5 Model-based Coding Basic Linear Prediction, LPC Mixed Excitation LPC (MELP) 6 Acknowledgments

25 Probability Density of speech By setting x[n] x, the histogram of speech samples can be approximated by a gamma density: p X (x) = or by a simpler Laplacian density: ( ) 1/2 3 e 3 x 2σx 8πσ x x p X (x) = 1 e 3 x σx 2σx

26 Densities comparison

27 Outline 1 Introduction 2 Statistical Models 3 Scalar Quantization Max Quantizer Companding Adaptive quantization Differential and Residual quantization 4 Vector Quantization The k-means algorithm The LBG algorithm 5 Model-based Coding Basic Linear Prediction, LPC Mixed Excitation LPC (MELP) 6 Acknowledgments

28 Coding and Decoding

29 Fundamentals of Scalar Coding Let s quantize a single sample speech value, x[n] into M reconstruction or decision levels: ˆx[n] = ˆx i = Q(x[n]), x i 1 < x[n] x i with 1 i M and x k denotes the M decision levels with 0 k M. Assign a codeword in each reconstruction level. Collection of codewords makes a codebook. Using B-bit binary codebook we can represent each 2 B different quantization (reconstruction) levels. Bit rate, I, is defined as: I = Bf s

30 Fundamentals of Scalar Coding Let s quantize a single sample speech value, x[n] into M reconstruction or decision levels: ˆx[n] = ˆx i = Q(x[n]), x i 1 < x[n] x i with 1 i M and x k denotes the M decision levels with 0 k M. Assign a codeword in each reconstruction level. Collection of codewords makes a codebook. Using B-bit binary codebook we can represent each 2 B different quantization (reconstruction) levels. Bit rate, I, is defined as: I = Bf s

31 Fundamentals of Scalar Coding Let s quantize a single sample speech value, x[n] into M reconstruction or decision levels: ˆx[n] = ˆx i = Q(x[n]), x i 1 < x[n] x i with 1 i M and x k denotes the M decision levels with 0 k M. Assign a codeword in each reconstruction level. Collection of codewords makes a codebook. Using B-bit binary codebook we can represent each 2 B different quantization (reconstruction) levels. Bit rate, I, is defined as: I = Bf s

32 Fundamentals of Scalar Coding Let s quantize a single sample speech value, x[n] into M reconstruction or decision levels: ˆx[n] = ˆx i = Q(x[n]), x i 1 < x[n] x i with 1 i M and x k denotes the M decision levels with 0 k M. Assign a codeword in each reconstruction level. Collection of codewords makes a codebook. Using B-bit binary codebook we can represent each 2 B different quantization (reconstruction) levels. Bit rate, I, is defined as: I = Bf s

33 Uniform quantization x i x i 1 =, 1 i M ˆx i = x i +x i 1 2, 1 i M is referred to as uniform quantization step size. Example of a 2-bit uniform quantization:

34 Uniform quantization: designing decision regions Signal range: 4σ x x[n] 4σ x Assuming B-bit binary codebook, we get 2 B quantization (reconstruction) levels Quantization step size, : and quantization noise. = 2x max 2 B

35 Classes of Quantization noise There are two classes of quantization noise: Granular Distortion: ˆx[n] = x[n] + e[n] where e[n] is the quantization noise, with: 2 e[n] 2 Overload Distortion: clipped samples

36 Assumptions Quantization noise is an ergodic white-noise random process: r e [m] = E(e[n]e[n + m]) = σ 2 e, m = 0 = 0, m 0 Quantization noise and input signal are uncorrelated: E(x[n]e[n + m]) = 0 m Quantization noise is uniform over the quantization interval p e (e) = 1, 2 e 2 = 0, otherwise

37 Assumptions Quantization noise is an ergodic white-noise random process: r e [m] = E(e[n]e[n + m]) = σ 2 e, m = 0 = 0, m 0 Quantization noise and input signal are uncorrelated: E(x[n]e[n + m]) = 0 m Quantization noise is uniform over the quantization interval p e (e) = 1, 2 e 2 = 0, otherwise

38 Assumptions Quantization noise is an ergodic white-noise random process: r e [m] = E(e[n]e[n + m]) = σ 2 e, m = 0 = 0, m 0 Quantization noise and input signal are uncorrelated: E(x[n]e[n + m]) = 0 m Quantization noise is uniform over the quantization interval p e (e) = 1, 2 e 2 = 0, otherwise

39 Dithering Definition (Dithering) We can force e[n] to be white and uncorrelated with x[n] by adding noise to x[n] before quantization!

40 Signal-to-Noise Ratio To quantify the severity of the quantization noise, we define the Signal-to-Noise Ratio (SNR) as: SNR = = σx 2 σ e 2 E(x 2 [n]) E(e 2 [n]) 1 N 1 N 1 N n=0 x2 [n] N 1 n=0 e2 [n] For uniform pdf and quantizer range 2x max : σe 2 = 2 12 x = max B Or and in db: and since x max = 4σ x : SNR = 3 22B ( xmax σx )2 ( ) xmax SNR(dB) 6B log 10 SNR(dB) 6B 7.2 σ x

41 Signal-to-Noise Ratio To quantify the severity of the quantization noise, we define the Signal-to-Noise Ratio (SNR) as: SNR = = σx 2 σ e 2 E(x 2 [n]) E(e 2 [n]) 1 N 1 N 1 N n=0 x2 [n] N 1 n=0 e2 [n] For uniform pdf and quantizer range 2x max : σe 2 = 2 12 x = max B Or and in db: and since x max = 4σ x : SNR = 3 22B ( xmax σx )2 ( ) xmax SNR(dB) 6B log 10 SNR(dB) 6B 7.2 σ x

42 Signal-to-Noise Ratio To quantify the severity of the quantization noise, we define the Signal-to-Noise Ratio (SNR) as: SNR = = σx 2 σ e 2 E(x 2 [n]) E(e 2 [n]) 1 N 1 N 1 N n=0 x2 [n] N 1 n=0 e2 [n] For uniform pdf and quantizer range 2x max : σe 2 = 2 12 x = max B Or and in db: and since x max = 4σ x : SNR = 3 22B ( xmax σx )2 ( ) xmax SNR(dB) 6B log 10 SNR(dB) 6B 7.2 σ x

43 Signal-to-Noise Ratio To quantify the severity of the quantization noise, we define the Signal-to-Noise Ratio (SNR) as: SNR = = σx 2 σ e 2 E(x 2 [n]) E(e 2 [n]) 1 N 1 N 1 N n=0 x2 [n] N 1 n=0 e2 [n] For uniform pdf and quantizer range 2x max : σe 2 = 2 12 x = max B Or and in db: and since x max = 4σ x : SNR = 3 22B ( xmax σx )2 ( ) xmax SNR(dB) 6B log 10 SNR(dB) 6B 7.2 σ x

44 Signal-to-Noise Ratio To quantify the severity of the quantization noise, we define the Signal-to-Noise Ratio (SNR) as: SNR = = σx 2 σ e 2 E(x 2 [n]) E(e 2 [n]) 1 N 1 N 1 N n=0 x2 [n] N 1 n=0 e2 [n] For uniform pdf and quantizer range 2x max : σe 2 = 2 12 x = max B Or and in db: and since x max = 4σ x : SNR = 3 22B ( xmax σx )2 ( ) xmax SNR(dB) 6B log 10 SNR(dB) 6B 7.2 σ x

45 Pulse Code Modulation, PCM B bits of information per sample are transmitted as a codeword instantaneous coding not signal-specific 11 bits are required for toll quality what is the rate for f s = 10kHz? For CD, with f s = and B = 16 (16-bit PCM), what is the SNR?

46 Pulse Code Modulation, PCM B bits of information per sample are transmitted as a codeword instantaneous coding not signal-specific 11 bits are required for toll quality what is the rate for f s = 10kHz? For CD, with f s = and B = 16 (16-bit PCM), what is the SNR?

47 Pulse Code Modulation, PCM B bits of information per sample are transmitted as a codeword instantaneous coding not signal-specific 11 bits are required for toll quality what is the rate for f s = 10kHz? For CD, with f s = and B = 16 (16-bit PCM), what is the SNR?

48 Pulse Code Modulation, PCM B bits of information per sample are transmitted as a codeword instantaneous coding not signal-specific 11 bits are required for toll quality what is the rate for f s = 10kHz? For CD, with f s = and B = 16 (16-bit PCM), what is the SNR?

49 Pulse Code Modulation, PCM B bits of information per sample are transmitted as a codeword instantaneous coding not signal-specific 11 bits are required for toll quality what is the rate for f s = 10kHz? For CD, with f s = and B = 16 (16-bit PCM), what is the SNR?

50 Pulse Code Modulation, PCM B bits of information per sample are transmitted as a codeword instantaneous coding not signal-specific 11 bits are required for toll quality what is the rate for f s = 10kHz? For CD, with f s = and B = 16 (16-bit PCM), what is the SNR?

51 Optimal decision and reconstruction level if x[n] p x (x) we determine the optimal decision level, x i and the reconstruction level, ˆx, by minimizing: D = E[(ˆx x) 2 ] = p x(x)(ˆx x) 2 dx and assuming M reconstruction levels ˆx = Q[x]: D = M = i=1 xi x i 1 p x (x)(ˆx i x) 2 dx So: D ˆx k = 0, 1 k M D x k = 0, 1 k M 1

52 Optimal decision and reconstruction level if x[n] p x (x) we determine the optimal decision level, x i and the reconstruction level, ˆx, by minimizing: D = E[(ˆx x) 2 ] = p x(x)(ˆx x) 2 dx and assuming M reconstruction levels ˆx = Q[x]: D = M = i=1 xi x i 1 p x (x)(ˆx i x) 2 dx So: D ˆx k = 0, 1 k M D x k = 0, 1 k M 1

53 Optimal decision and reconstruction level if x[n] p x (x) we determine the optimal decision level, x i and the reconstruction level, ˆx, by minimizing: D = E[(ˆx x) 2 ] = p x(x)(ˆx x) 2 dx and assuming M reconstruction levels ˆx = Q[x]: D = M = i=1 xi x i 1 p x (x)(ˆx i x) 2 dx So: D ˆx k = 0, 1 k M D x k = 0, 1 k M 1

54 Optimal decision and reconstruction level, cont. The minimization of D over decision level, x k, gives: x k = ˆx k+1 + ˆx k, 1 k M 1 2 The minimization of D over reconstruction level, ˆx k, gives: ] xdx ˆx k = x k x k 1 [ p x (x) xk x p x (s)ds k 1 = x k x k 1 p x (x)xdx

55 Optimal decision and reconstruction level, cont. The minimization of D over decision level, x k, gives: x k = ˆx k+1 + ˆx k, 1 k M 1 2 The minimization of D over reconstruction level, ˆx k, gives: ] xdx ˆx k = x k x k 1 [ p x (x) xk x p x (s)ds k 1 = x k x k 1 p x (x)xdx

56 Example with Laplacian pdf

57 Principle of Companding

58 Companding examples Companding examples: Transformation to a uniform density: g[n] = T (x[n]) = x[n] p x(s)ds 1 2, 1 2 g[n] 1 2 = 0 elsewhere µ-law: x max ) T (x[n]) = x max log (1 + µ x[n] log (1 + µ) sign(x[n])

59 Companding examples Companding examples: Transformation to a uniform density: g[n] = T (x[n]) = x[n] p x(s)ds 1 2, 1 2 g[n] 1 2 = 0 elsewhere µ-law: x max ) T (x[n]) = x max log (1 + µ x[n] log (1 + µ) sign(x[n])

60 Adaptive quantization

61 Differential and Residual quantization

62 Outline 1 Introduction 2 Statistical Models 3 Scalar Quantization Max Quantizer Companding Adaptive quantization Differential and Residual quantization 4 Vector Quantization The k-means algorithm The LBG algorithm 5 Model-based Coding Basic Linear Prediction, LPC Mixed Excitation LPC (MELP) 6 Acknowledgments

63 Motivation for VQ

64 Comparing scalar and vector quantization

65 Distortion in VQ Here we have a multidimensional pdf p x (x): D = E[(ˆx x) T (ˆx x)] = (ˆx x)t (ˆx x)p x (x)dx = M i=1 x C i (r i x) T (r i x)p x (x)dx Two constraints: A vector x must be quantized to a reconstruction level r i that gives the smallest distortion: C i = {x : x r i 2 x r l 2, l = 1, 2,, M} Each reconstruction level r i must be the centroid of the corresponding decision region, i.e., of the cell C i : x r i = m C i x m i = 1, 2,, M x m C i 1

66 Distortion in VQ Here we have a multidimensional pdf p x (x): D = E[(ˆx x) T (ˆx x)] = (ˆx x)t (ˆx x)p x (x)dx = M i=1 x C i (r i x) T (r i x)p x (x)dx Two constraints: A vector x must be quantized to a reconstruction level r i that gives the smallest distortion: C i = {x : x r i 2 x r l 2, l = 1, 2,, M} Each reconstruction level r i must be the centroid of the corresponding decision region, i.e., of the cell C i : x r i = m C i x m i = 1, 2,, M x m C i 1

67 The k-means algorithm S1: D = 1 N 1 (ˆx k x k ) T (ˆx k x k ) N k=0 S2: Pick an initial guess at the reconstruction levels {r i } S3: For each x k elect an r i closest to x k. Form clusters (clustering step) S4: Find the mean of x k in each cluster which gives a new r i. Compute D. S5: Stop when the change in D over two consecutive iterations is insignificant.

68 The LBG algorithm Set the desired number of cells: M = 2 B Set an initial codebook C (0) with one codevector which is set as the average of the entire training sequence, x k, k = 1, 2,, N. Split the codevector into two and get an initial new codebook C (1). Perform a k-means algorithm to optimize the codebook and get the final C (1) Split the final codevectors into four and repeat the above process until the desired number of cells is reached.

69 The LBG algorithm Set the desired number of cells: M = 2 B Set an initial codebook C (0) with one codevector which is set as the average of the entire training sequence, x k, k = 1, 2,, N. Split the codevector into two and get an initial new codebook C (1). Perform a k-means algorithm to optimize the codebook and get the final C (1) Split the final codevectors into four and repeat the above process until the desired number of cells is reached.

70 The LBG algorithm Set the desired number of cells: M = 2 B Set an initial codebook C (0) with one codevector which is set as the average of the entire training sequence, x k, k = 1, 2,, N. Split the codevector into two and get an initial new codebook C (1). Perform a k-means algorithm to optimize the codebook and get the final C (1) Split the final codevectors into four and repeat the above process until the desired number of cells is reached.

71 The LBG algorithm Set the desired number of cells: M = 2 B Set an initial codebook C (0) with one codevector which is set as the average of the entire training sequence, x k, k = 1, 2,, N. Split the codevector into two and get an initial new codebook C (1). Perform a k-means algorithm to optimize the codebook and get the final C (1) Split the final codevectors into four and repeat the above process until the desired number of cells is reached.

72 The LBG algorithm Set the desired number of cells: M = 2 B Set an initial codebook C (0) with one codevector which is set as the average of the entire training sequence, x k, k = 1, 2,, N. Split the codevector into two and get an initial new codebook C (1). Perform a k-means algorithm to optimize the codebook and get the final C (1) Split the final codevectors into four and repeat the above process until the desired number of cells is reached.

73 Outline 1 Introduction 2 Statistical Models 3 Scalar Quantization Max Quantizer Companding Adaptive quantization Differential and Residual quantization 4 Vector Quantization The k-means algorithm The LBG algorithm 5 Model-based Coding Basic Linear Prediction, LPC Mixed Excitation LPC (MELP) 6 Acknowledgments

74 Basic coding scheme in LPC Vocal tract system function: where H(z) = P(z) = A 1 P(z) p a k z 1 k=1 Input is binary: impulse/noise excitation. If frame rate is 100 frames/s and we use 13 parameters (p = 10, 1 for Gain, 1 for pitch, 1 for voicing decision) we need 1300 parameters/s, instead of 8000 samples/s for f s = 8000Hz.

75 Basic coding scheme in LPC Vocal tract system function: where H(z) = P(z) = A 1 P(z) p a k z 1 k=1 Input is binary: impulse/noise excitation. If frame rate is 100 frames/s and we use 13 parameters (p = 10, 1 for Gain, 1 for pitch, 1 for voicing decision) we need 1300 parameters/s, instead of 8000 samples/s for f s = 8000Hz.

76 Basic coding scheme in LPC Vocal tract system function: where H(z) = P(z) = A 1 P(z) p a k z 1 k=1 Input is binary: impulse/noise excitation. If frame rate is 100 frames/s and we use 13 parameters (p = 10, 1 for Gain, 1 for pitch, 1 for voicing decision) we need 1300 parameters/s, instead of 8000 samples/s for f s = 8000Hz.

77 Scalar quantization within LPC For 7200 bps: Voiced/unvoiced decision: 1 bit Pitch (if voiced): 6 bits (uniform) Gain: 5 bits (nonuniform) Poles d i : 10 bits (nonuniform) [5 bits for frequency and 5 bits for bandwidth] 6 poles = 60 bits So: ( ) 100 frames/s = 7200 bps

78 Scalar quantization within LPC For 7200 bps: Voiced/unvoiced decision: 1 bit Pitch (if voiced): 6 bits (uniform) Gain: 5 bits (nonuniform) Poles d i : 10 bits (nonuniform) [5 bits for frequency and 5 bits for bandwidth] 6 poles = 60 bits So: ( ) 100 frames/s = 7200 bps

79 Refinements to the basic LPC coding scheme Companding in the form of a logarithmic operator on pitch and gain Instead of poles use the reflection (or the PARCOR) coefficients k i,(nonuniform) Companding of k i : g i = T [k ( i ] ) = log 1 ki 1+k i Coefficients g i can be coded at 5-6 bits each! (which results in 4800 bps for an order 6 predictor, and 100 frames/s) Reduce the frame rate by a factor of two (50 frames/s) gives us a bit rate of 2400 bps

80 Refinements to the basic LPC coding scheme Companding in the form of a logarithmic operator on pitch and gain Instead of poles use the reflection (or the PARCOR) coefficients k i,(nonuniform) Companding of k i : g i = T [k ( i ] ) = log 1 ki 1+k i Coefficients g i can be coded at 5-6 bits each! (which results in 4800 bps for an order 6 predictor, and 100 frames/s) Reduce the frame rate by a factor of two (50 frames/s) gives us a bit rate of 2400 bps

81 Refinements to the basic LPC coding scheme Companding in the form of a logarithmic operator on pitch and gain Instead of poles use the reflection (or the PARCOR) coefficients k i,(nonuniform) Companding of k i : g i = T [k ( i ] ) = log 1 ki 1+k i Coefficients g i can be coded at 5-6 bits each! (which results in 4800 bps for an order 6 predictor, and 100 frames/s) Reduce the frame rate by a factor of two (50 frames/s) gives us a bit rate of 2400 bps

82 Refinements to the basic LPC coding scheme Companding in the form of a logarithmic operator on pitch and gain Instead of poles use the reflection (or the PARCOR) coefficients k i,(nonuniform) Companding of k i : g i = T [k ( i ] ) = log 1 ki 1+k i Coefficients g i can be coded at 5-6 bits each! (which results in 4800 bps for an order 6 predictor, and 100 frames/s) Reduce the frame rate by a factor of two (50 frames/s) gives us a bit rate of 2400 bps

83 Refinements to the basic LPC coding scheme Companding in the form of a logarithmic operator on pitch and gain Instead of poles use the reflection (or the PARCOR) coefficients k i,(nonuniform) Companding of k i : g i = T [k ( i ] ) = log 1 ki 1+k i Coefficients g i can be coded at 5-6 bits each! (which results in 4800 bps for an order 6 predictor, and 100 frames/s) Reduce the frame rate by a factor of two (50 frames/s) gives us a bit rate of 2400 bps

84 VQ in LPC coding A 10-bit codebook (1024 codewords), 800 bps VQ provides a comparable quality to a 2400 bps scalar quantizer. A 22-bit codebook ( codewords), 2400 bps VQ provides a higher output speech quality.

85 Unique components of MELP Mixed pulse and noise excitation Periodic or aperiodic pulses Adaptive spectral enhancements Pulse dispersion filter

86 Line Spectral Frequencies (LSFs) in MELP LSFs for a pth order all-pole model are defines as follows: 1 Form two polynomials: P(z) = A(z) + z (p+1) A(z 1 ) Q(z) = A(z) z (p+1) A(z 1 ) 2 Find the roots of P(z) and Q(z), ωi which are on the unit circle. 3 Exclude trivial roots at ωi = 0 and ω i = π.

87 MELP coding For a 2400 bps: 34 bits allocated to scalar quantization of the LSFs 8 bits for gain 7 bits for pitch and overall voicing 5 bits for multi-band voicing 1 bit for the jittery state which is 54 bits. With a frame rate of 22.5 ms, we get an 2400 bps coder.

88 Outline 1 Introduction 2 Statistical Models 3 Scalar Quantization Max Quantizer Companding Adaptive quantization Differential and Residual quantization 4 Vector Quantization The k-means algorithm The LBG algorithm 5 Model-based Coding Basic Linear Prediction, LPC Mixed Excitation LPC (MELP) 6 Acknowledgments

89 Acknowledgments Most, if not all, figures in this lecture are coming from the book: T. F. Quatieri: Discrete-Time Speech Signal Processing, principles and practice 2002, Prentice Hall and have been used after permission from Prentice Hall

1. Probability density function for speech samples. Gamma. Laplacian. 2. Coding paradigms. =(2X max /2 B ) for a B-bit quantizer Δ Δ Δ Δ Δ

1. Probability density function for speech samples. Gamma. Laplacian. 2. Coding paradigms. =(2X max /2 B ) for a B-bit quantizer Δ Δ Δ Δ Δ Digital Speech Processing Lecture 16 Speech Coding Methods Based on Speech Waveform Representations and Speech Models Adaptive and Differential Coding 1 Speech Waveform Coding-Summary of Part 1 1. Probability