Some Notes on the Psychoacoustics and Signal Processing of RASTA-PLP - - PowerPoint PPT Presentation

Some Notes on the Psychoacoustics and Signal Processing of RASTA-PLP Analysis of Speech

• Introduction
• Reflection Masking Model
• RASTA-PLP Processing
• RASTA applied to RIR filtered speech

Wire Communication Laboratory - University of Patras

Jörg Buchholz

Comments on Perceptual Modelling

Psychoacoustics Physiology Perceptual Model (Masking Model) Speech enhancement in reverberant environments

Comparison Description Application Requirements Requirements

Illustration of Masking and Suppression

masker suppressor suppressor test signals test signals post (forward) masking pre (backward) masking simultaneous masking time frequency

Structure of the Masking Module

TMM Simultaneous Masking Directivity Module s (t)

i

s (t)

i-1

s (t)

i+1

Module Two-Tone Suppression Module (BP-Filterbank) Transformation / Resythesis Feature Vectors / Audible Signal TMM TMM TMM

Block Diagramm of the RASTA-PLP Method

Speech compressing static NL (log) CB-integration (mel-scale) FFT expanding static NL (Exp) linear BP-filtering equal loudness curve IFFT / IDFT power law of hearing cepstral recursion solving set of linear equations cepstral coefficients of RASTA-PLP model RASTA

Time / Frequency Analysis

time / ms 10 50 40 60 20 30 frequency frequency amplitude 1 fa/2 fa/2 power spectrum mel-scale power spectrum short time analysis i+3 i+4 i+1 i+2 i+5 i time trajectory k FFT CB integration

CB-integration (mel-scale)

10

2

10

3

10

4

• 40
• 20

20 Magnitude in dB 10

2

10

3

10

4

• 60
• 40
• 20

20 40 frequency / Hz Magnitude in dB

Overview of some RASTA Methods

• Additive Noise (uncorrelated) → Lin-RASTA
• Convolutional Noise → Log-RASTA
• Convolutional and Additive Noise → J-RASTA (Lin/Log)

Optimal J depending on the noise power!

( ) ( ) ( ) ( ) ( )

( )

( )

( )

s t h t S H S H

FT

⊗ → ⋅ → + ω ω ω ω

log

log log

y J x x e J e J

y y

= + ⋅ = − ≈ ln( ) 1 1

( ) ( ) ( ) ( )

s t n t S N

FT

+ → + ω ω

Rasta BP-Filter

50 100 150 200 250 300 350 400 0.1 0.2 0.3 Time / ms Amplitude 10

• 2

10

• 1

10 10

1

10

2

• 40
• 30
• 20
• 10

10 Modulation Frequency / Hz Magnitude in dB

( )

H z z z z z z = ⋅ ⋅ + − − −

− − − −

01 2 2 1 0 98

4 1 3 4 1

. .

LIN-RASTA-Processing of clean speech

500 1000 1500 2000 0.5 1 500 1000 1500 2000

• 0.5

0.5 1 normalis ed amplitude 500 1000 1500 2000 0.5 1 time / ms

Time Trajectory (1 kHz band) BP-filtered Time Trajectory Negative values set to zero

Lin-RASTA-Processing of noisy speech (-5 dB SNR)

500 1000 1500 2000 0.5 1 500 1000 1500 2000

• 0.5

0.5 1 normalis ed amplitude 500 1000 1500 2000 0.5 1 time / ms

Time Trajectory (1 kHz band) BP-filtered Time Trajectory Negative values set to zero

RASTA applied to Room Impulse Responses

Basic Assumptions for RASTA processing:

• Analysis Window length >> Filter (RIR) length
• Filter (RIR) should be constant (slowly changing) for

a duration >> Analysis Window length

• For ASR: Window length ≈ 20-30 ms

Conflict!!!

Baseline Multiresolution Clean 8,6 % 13,5 % Reverberant 34,8 % 22,8 % Isolated Digits Word Error Rates

Multiresolution Processing Concept (Avendano)

x(n) n x(n) n X(n , )

2 2k

ω X(n , )

1 1k

ω n2 ω1 A

• ✁

n1 ω2

• 12 dB

0 dB 0 dB

• 6 dB

+18 dB

Equal-Loudness Curves

Loudness function (Zwicker)