Frequency Domain Coding of Speech 主講人:虞台文
Content Introduction The Short-Time Fourier Transform The Short-Time Discrete Fourier Transform Wide-Band Analysis/Synthesis Sub-Band Coding
Frequency Domain Coding of Speech Introduction
Speech Coders Waveform Coders – Attempt to reproducing the original waveform according to some fidelity criteria – Performance: successful at producing good quality, robust speech. Vocoders – Correlated with speech production model. – Performance: more fragile and more model dependent. – Lower bit rate
Frequency-Domain Coders Sub-band coder (SCB). Adaptive Transform Coding (ATC). Multi-band Excited Vocoder (MBEV). Noise Shaping in Speech Coders.
Classification of Speech Coders
Frequency Domain Coding of Speech The Short-Time Fourier Transform
Definition of STFT Interpretations: Filter Bank Interpretation Block Transform Interpretation
Filter Bank Interpretation is fixed at 0. f ( m ) Analysis Filter
Filter Bank Interpretation h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) x(n)x(n)
Modulation 00
Filter Bank Interpretation 00 Lowpass Filter Modulation
Filter Bank Interpretation h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) x(n)x(n) Modulated Subband signals
Block Transform Interpretation n is fixed at n 0. Windowed Data Analysis Window FT of Windowed Data
Block Transform Interpretation n is fixed at n 0. n1n1 n2n2 n3n nrnr
Analysis/Synthesis Equations Analysis Synthesis In what condition we will have
Analysis/Synthesis Equations Analysis Synthesis Replace r with n+r
Analysis/Synthesis Equations Analysis Synthesis Therefore, if
Analysis/Synthesis Equations More general, Analysis Synthesis Therefore, if
Examples
h (0) x ( n )
Examples
Frequency Domain Coding of Speech The Short-Time Discrete Fourier Transform
Definition of STDFT Analysis: Synthesis: In what condition we will have
Synthesis 1
We need only one period. Therefore, the condition is respecified as:
Implementation Consideration n Frequency k 0 Spectrogram
Sampling n Frequency k 0 Spectrogram R2R2R3R3R4R4R
Sampled STDFT Analysis: Synthesis: In what condition we will have
Sampled STDFT Analysis: Synthesis: In what condition we will have
Frequency Domain Coding of Speech Wide-Band Analysis/Synthesis
Short-Time Synthesis --- Filter Bank Summation STFT h(n)h(n) h(n)h(n) x(n)x(n) Lowpass Filter
Short-Time Synthesis --- Filter Bank Summation STFT
Short-Time Synthesis --- Filter Bank Summation |H(e j )| |H k (e j )| kk Lowpass filterBandpass filter
Short-Time Synthesis --- Filter Bank Summation hk(n)hk(n) hk(n)hk(n) x(n)x(n) Bandpass Filter h(n)h(n) h(n)h(n) x(n)x(n) Lowpass Filter Lowpass representation of for the signal in a band centered at k.
Short-Time Synthesis --- Filter Bank Summation hk(n)hk(n) hk(n)hk(n) x(n)x(n) Bandpass Filter h(n)h(n) h(n)h(n) x(n)x(n) Lowpass Filter Encoding one bandDecoding one band
Short-Time Synthesis --- Filter Bank Summation h1(n)h1(n) h1(n)h1(n) x(n)x(n) h0(n)h0(n) h0(n)h0(n) hN1(n)hN1(n) hN1(n)hN1(n) Analysis Synthesis
Short-Time Synthesis --- Filter Bank Summation h1(n)h1(n) h1(n)h1(n) x(n)x(n) h0(n)h0(n) h0(n)h0(n) hN1(n)hN1(n) hN1(n)hN1(n) Analysis Synthesis
Short-Time Synthesis --- Filter Bank Summation h1(n)h1(n) h1(n)h1(n) x(n)x(n) h0(n)h0(n) h0(n)h0(n) hN1(n)hN1(n) hN1(n)hN1(n) Analysis Synthesis
Equal Spaced Ideal Filters 11 22 33 44 55 22 1 0 N = 6
Equal Spaced Ideal Filters h1(n)h1(n) x(n)x(n) h0(n)h0(n) hN1(n)hN1(n) What condition should be satisfied so that y(n)=x(n)?
Equal Spaced Ideal Filters Equal spaced sampling of H ( e j ) Inverse discrete FT of H ( e j ) Time-Aliased version of h ( n )
Equal Spaced Ideal Filters Consider FIR, i.e., h(n) is of duration of L samples. 0 L1L1 n h(n)h(n) In case that N L,
Equal Spaced Ideal Filters
h1(n)h1(n) h1(n)h1(n) x(n)x(n) h0(n)h0(n) h0(n)h0(n) hN1(n)hN1(n) hN1(n)hN1(n) L1L1 n h(n)h(n) x(n) can always be Reconstructed if N L,
Equal Spaced Ideal Filters h1(n)h1(n) h1(n)h1(n) x(n)x(n) h0(n)h0(n) h0(n)h0(n) hN1(n)hN1(n) hN1(n)hN1(n) L1L1 n h(n)h(n) x(n) can always be Reconstructed if N L, Does x(n) can still be reconstructed if N<L? If affirmative, what condition should be satisfied?
Equal Spaced Ideal Filters h1(n)h1(n) h1(n)h1(n) x(n)x(n) h0(n)h0(n) h0(n)h0(n) hN1(n)hN1(n) hN1(n)hN1(n) p(n)p(n)
p(n)p(n) Signal can be reconstructed If it equals to ( n m ).
Typical Sequences of h (n) Ideal lowpass filter with cutoff at /N. 0 NN 2N2N N2N2N3N3N4N4N p(n)p(n) N 0 NN 2N2N N2N2N3N3N4N4N h(n)h(n) 1/N
Typical Sequences of h (n) 0 NN 2N2N N2N2N3N3N4N4N p(n)p(n) N 0 NN 2N2N N2N2N3N3N4N4N h(n)h(n) h(0) LL 2L2L L2L2L3L3L4L4L N LN L
Typical Sequences of h (n) 0 NN 2N2N N2N2N3N3N4N4N p(n)p(n) N 0 NN 2N2N N2N2N3N3N4N4N h(n)h(n) h(0) 1/N A causal FIR lowpass filter
Typical Sequences of h (n) 0 NN 2N2N N2N2N3N3N4N4N p(n)p(n) N 0 NN 2N2N N2N2N3N3N4N4N h(n)h(n) h(0) 1/N A causal IIR lowpass filter
Filter Back Implementation for a Single Channel hk(n)hk(n) x(n)x(n) h(n)h(n) x(n)x(n) Analysis Synthesis
hk(n)hk(n) x(n)x(n) h(n)h(n) x(n)x(n) Filter Back Implementation for a Single Channel R:1 1:R Analysis Synthesis Decimator Interpolator
hk(n)hk(n) x(n)x(n) h(n)h(n) x(n)x(n) Filter Back Implementation for a Single Channel R:1 1:R Analysis Synthesis Decimator Interpolator Depends on the bandwidth of h(n). R=?
Frequency Domain Coding of Speech Sub-Band Coding
Analysis Synthesis Filter Bank Implementation (Direct Implementation) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) x(n)x(n) R:1 1:R f(n)f(n) f(n)f(n) f(n)f(n) f(n)f(n) f(n)f(n) f(n)f(n) f(n)f(n) f(n)f(n) x(n)x(n) Complex Channels R=2 B Bandwidth B/2
Filter Bank Implementation (Practical Implementation) 0 B kk 0 B k 0 B/2 B/2 0 B/2 B/2 0 BB 0 B 0 BB B
Filter Bank Implementation (Practical Implementation) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) x(n)x(n)......
h(n)h(n) h(n)h(n) x(n)x(n) h(n)h(n) h(n)h(n)
h(n)h(n) h(n)h(n) x(n)x(n) h(n)h(n) h(n)h(n) D:1 Why?
Filter Bank Implementation (Practical Implementation) h(n)h(n) h(n)h(n) x(n)x(n) h(n)h(n) h(n)h(n) D:1
h(n)h(n) h(n)h(n) x(n)x(n) h(n)h(n) h(n)h(n) Filter Bank Implementation (Practical Implementation)
x(n)x(n) h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) D:1 2D:1
Filter Bank Implementation (Practical Implementation) ADPCM CODEC h(n)h(n) h(n)h(n) h(n)h(n) h(n)h(n) 2D:1 f(n)f(n) f(n)f(n) f(n)f(n) f(n)f(n) Filter Bank Analysis Filter Bank Analysis Sub-Band Coder Modification Sub-Band Coder Modification Filter Bank Synthesis Filter Bank Synthesis