Voice Quality Enhancement

1
Voice Quality Enhancement

• Shi Chao Zhang, Kevin McCook,
• Kurtis Chang
• T.A. Matt Olson

2
Objective

• Build a hands-free speakerphone system
• Apply Digital Signal Processing
• Remove acoustic echo from transmission
• Adaptive Echo Cancellation, Echo Suppression
• Reduce noise
• Frequency Domain Spectral Subtraction

3
Design Review
Incoming Signal
Audio Amp.
Computer 1
Input 1
Spectral Subtraction
Speaker
DSP
Output
DSP
Input 2
Computer 2
Microphone Preamp.
mic
Outgoing Signals
4
Digital Signal Processing

• Texas Instruments TMS320C54X DSP
• Sampling rate 44.1 kHz
• Frequency response 0 22.05 kHz
• Voice Signals Are Bandlimited
• 300 to 3400 Hz
• Implement multi-rate processing
• Increases available processing time
• Improves efficiency of algorithms.

5
Decimation/Interpolation

• Changes Sampling Rate of DSP

Processing
6
Echo Canceling

• Input 1 Incoming Signal
• Input 2 Signal received by microphone

?
AFIR
LPF
D ?
Input 1

LPF
D ?
Input 2
Output
7
Adaptive Filter Based E.C.

• Filter converges to transfer function
• Uses LMS algorithm to update coefficients
• Uses error to calculate
• Parameter Step Size affects convergence rate
• Removes Unwanted Signal Digitally

Unknown Transfer Function
8
Adaptive Filter Design

• Had to choose a step size
• 0.06 was ideal step size
• Had to determine length
• Longer Adaptive Filter could remove longer
  lasting echos
• 510 was longest possible length

9
Noise Reduction

• Conceptual overview
• Determination of noise for each frequency
• Subtraction of lowest noise in time period from
  each frequency

10
Spectral Subtraction

• Algorithmic overview

11
Spectral Subtraction

• MatLab simulation results

Sample 1
Sample 2
12
Spectral Subtraction

• DSP simulation
• FFT and IFFT reduce precision
• 1024-point FFT-IFFT pair

Sample DSP output
13
Echo Suppression Overview

• eliminates near-end signal when no double-talk is
  detected
• uses voice-activity detector to determine
  double-talk state
• uses comfort noise generation for when there is
  no near-end signal

14
Voice-Activity Detector

• -uses recursive power algorithm to compare power
  level vs. double-talk
• threshold
• -P(i) (1-alpha) P(i-1) alpha input(i)
• where alpha time constant
• -uses holdover threshold to prevent clipping of
  speech signals

15
comfort noise generator

• uses a pseudo-noise generator to produce the
  'random' bits
• -using a shift register of size N, bits will
  repeat every 2N - 1 times

16
echo suppression initial simulations

• through research, decided to implement recursive
  power algorithm
• used matlab to produce some preliminary test
  results and then adjust variables

17
echo suppression dsp implementation and testing

• after matlab testing, implemented voice activity
  detector in C and
• comfort noise generation in assembly
• then, had to adjust all variables to perform well
  with our speakerphone
• overall, the echo suppression worked successfully

18
Functional Tests

• Made recordings of unprocessed signal and
  processed signal simultaneously, then compared
  outputs
• Matched the gain of the microphone and speaker
  signals

19
Circuit Diagram

• Microphone Preamp

Used integrated circuit for audio amp Motorola
MC34119P
20
Successes

• Substantial reduction of echo noise through the
  use of an adaptive filter
• Successful determination of non-speech periods
  through echo suppression

21
Challenges

• A suitable rate of change for adaptive
  coefficients has to be determined ahead of time
• The number of coefficients is limited by the
  processing power of the DSP
• Noise reduction (spectral subtraction) was
  impractical to implement on the DSP due to
  reduction of precision in the FFT algorithm used

22
Recommendations

• Use Infinite Impulse Response Filters for
  anti-aliasing
• Uses less memory
• Use adaptive filter for comfort noise generator
• Scales noise added to the value of noise
  calculated in spectral subtraction

23
Conclusion

• The majority of the project worked at the end
• Had problems converting matlab code to assembly
  for Noise Reduction

24

• THANK YOU!!