4. In time domain, the
strength of sound
decays as the time
goes by.
C
B7
5. In frequency domain, a chord
has its fundamental frequency
and integer multiple of
fundamental frequency.
Different musical instruments
has different weights of
fundamental frequency and
integer multiple frequencies.
Timbers are discriminated these
combinations.
6. Part 2: Different
Approaches
Data gathering and format
Tool: Wavepad. Record a chord. And save it in WAV format in 2sec.
Matlab read WAV file and generate a 1xn matrix, each number in the
matrix represents the sound’s strength in corresponding time.
8. Why use…
-Band pass filter: guitar produce sound frequency
between ~15Hz - ~5000Hz.
-Guassian Smoothing: Required because we need
tolerance to the existance of guitar tuning error,
measurement error, computational error.
10. Approach 1: Mixture Component analysis
L*C = test
- L is formed by the 10000x1 chord feature vectors
of different chord.
We used 8 chords: A B7 C D E F G G7.
L = [A; B7; C; D; E; F; G; G7]; (10000x8).
- C is the coefficient matrix. (8x1)
- test is chord feature vector to be
tested(10000x1).
From equation:
Test is mixed by L with different percentage (c).
11. Approach 1: Mixture Component analysis
Least square solution:
C = inv(L'*L)*L'*test
We choose the biggest c_m.
Quality factor Q = c_m/sum(abs(C)) .
When Q > threshold, test data is
one of the chord in our database.
12. Approach 2: Principal Component Analysis
• Everything is same with eigenface analysis. Input is also
large dimension a vector.
13. Compare
Principal Component Analysis
With
Mixture Component Analysis:
In our content,
PCA does a better job in determining if test data is one of the
chord in our database.
MCA does better in recognition.
15. Drawback
Database is 1 sample for 1 chord, high error.
Remedy
Use LDA or multi dimension Guassian pdf.
(their database can be n samples for 1 chord).