Professional Documents
Culture Documents
SSP Sheets 2007 09
SSP Sheets 2007 09
SSP Sheets 2007 09
Filterbank Analysis
Bandlter analysis
Sound to MFCC
A Short-time FT to calculate the power (K spectral values Sk ) Filterbank with J triangular lters Hj Pj = k Hjk Sk for 1 j J where k Hjk = 1; j Cosine transform of the lterbank output cm = J1 cos m (j + 0.5) log Pj j=0 J Praat datatype MelFilter
Amplitude
0.5
0 0
8000
Amplitude
0.5
0 0
3000
Hertz to Mel
3000
From Hz to mel
mel = 2595 log(1 + Hertz/700)
Mel to Hertz
8000
Hertz
1000 mel
2000
3000
0.5
1.5
2.5
Frequency (mel)
8000 7000 Frequency (Hz) 6000 5000 4000 3000 2000 1000 0 0 0.5 1 1.5 2 2.5 3
train/dr1/mcpm0/sa1.wav To MelFilter... 0.025 ...0.005 100 100 0 To Spectrogram... 0.025 ...8000 0.005 20 Gaussian
Sound: To MelFilter...
Bandlter Analysis
Mel Frequency Cepstral Coecients References
40
20
0 Frequency (Hz)
8000
The bandltering process: Filterbank with J triangular lters Hj Pj = k Hjk Sk for 1 j J where k Hjk = 1; j
MelFilter: To MFCC...
References
R. Vergin & D. OShaughnessy (1999), Generalized Mel Frequency Cepstral Coecients for Large-Vocabulary Speaker-Independent Continuous-Speech Recognition, IEEE Trans. on Speech and Audio Processing 7, 525532.