Voice Activated Switch

1
Voice Activated Switch

• ECE345 Senior Design Project
• By
• Brad Griffis
• Sung Kim
• Ronald Roth
• November 30, 2001

2
Introduction

• Meet needs of handicapped UHS student
• Replace the original switch with our product
• Clip-on microphone with digital signal processing
• Switch-controlled power strip

3
Objectives

• Reliability for Matts voice-training
• Sensitivity for voice-activating

4
Overall Design
                                       
 
Lavalier MX185
SCM268 Mixer
DSK 5402
Power Supply Device
Electronic Device
5
Algorithm Development

• X referring to the spectrum of the observed
  speech signal
• E to the excitation component
• V to the vocal tract shaping of the excitation
  spectrum

6
DSP Flowchart - Input
7
DSP Flowchart - FFT
8
DSP Flowchart Log Magnitude
9
DSP Flowchart Cepstral Coeffs
10
Simulation on Cep. Coef.
11
Euclidean Distance

• Ci represents the ith cepstral coefficient of the
  windowed signal
• ?i represents the ith cepstral coefficient of the
  template

12
Simulation on Euclidean Dist.
13
Template (ah trimmed ah)
14
Percent Passing for ah, iy, oh
15
Percent Passing for shoe bench
16
Percent Ratio of ah to oh shoe bench
17
Threshold and Percent Passing

• Optimal percent passing chosen conservatively at
  0.30

18
Number of Cepstral Coeffs

• Trial and error
• 13 worked best

19
Ah Detection

• Three Ahs detected from ThreeAhs.wav at 1.61
  2.20 sec., 5.78 6.09 sec., and 8.35 8.94 sec.

20
Detecting Matts Ah

• One of Matts ahs
• detected by the
• Algorithm

21
Software Flowchart
C
Update Percent Passing Ah
Get Input
Ahs_state0
Ahs_state1
Depend on the cond.
Cepstrum( )
No Signal
Generate Sinusoidal Signal
Compare Cep. Coeffs. to template, getting
Euclidean Dist.
C

• I/O interrupt is called whenever the input buffer
  is full. After this, DSP continues its operation.

22
Bit Reverse
Replace oldest 10 ms samples with newest in
buffer
FFT
Normalize input to -1, 1
LogFFT
Replace the newest samples in buffer with the
normalized input
Bit Reverse
Hamming Window
IFFT
Zero Padding
Get first N Coeffs.
Convert to q15
23
A
No
Yes
Cond2
No
Cond1
Cond3
Cond4
Yes
No
No
Yes
Update ahs buffer
Yes
Ah_state 1
Ah_state 0
Increment ah
Update ahs buffer
A

• Cond1 Euclidean distance lt Threshold
  Cond3 ah_state 0
• Cond2 Percent_ahs gt Percent_ahs_Threshold
  Cond4 ah_state 1

24
Reverse-engineering

• Reverse-engineering of the original switch from
  enabling device Corp
• Output of the switch
• -When there is an input
• No output for 23 seconds
• -When there is no input
• 60 Hz 50 mVpp sinusoidal signal generated

25
DSP Board
26
Mixer Setup
1 should be plugged into standard outlet 3
should be set to mic 4 should be connected to
the DSP input 7 phantom power should be turned
on for a condenser mic (e.g. lavalier) 8 XLR
input from condenser mic (lavalier) plugs in here
27
I/O

• Codec
• Serial Port
• DMA
• Memory
• CPU
• Threading/interrupts

28
Verification

• Module testing
• Incremental coding

29
Challenges

• Memory Mapping
• Compiler Issues
• Performing an IFFT using an FFT

30
Successes

• Consistent cepstral results
• Determining optimal parameter values
• Getting Matlab to detect ahs in a file
• Reverse engineering power strip input
• Implementing I/O involving overlapping windows
• Calculation/Verification up to IFFT

31
Recommendations

• Use C5409 DSP Chip
• Case for DSP Board
• Detection of other phonemes

32
Acknowledgements

• Donation from Shure
• SCM268 mixer
• MX185 cardioid lavalier microphone
• Support from ECE department
• C5402 DSK board
• Prof. Hasegawa-Johnson