Beruflich Dokumente
Kultur Dokumente
Representa.ons
Visual Representa.ons of Sound Waves
.01 sec
200 Hz
.01 sec.
.005 sec
.003 sec
300 Hz
Power Spectrum.
A a descrip.on of frequency components that make up a sound.
Represented by a plot of the level of each of the pure-tone components as
a func.on of amplitude over frequency =Power Spectrum
High frequency right
Low frequency leI
Height of line amplitude
Waveform vs. Spectrum.
Waveform: Power Spectrum:
Advantage: provides Advantage: Provides
good temporal good frequency
resolu.on (dura.on, resolu.on (f0, harmonics
whole uLerance); f0; Disadvantage: Does not
changes in amplitude provide good temporal
over .me resolu.on (shows one
Disadvantage: Does not point in .me); formant
provide good frequency transi.ons
resolu.on; formant
transi.ons
Spectrography
Ver.cal stria.on = gloLal pulses
Darkness = intensity
dark horizontal bands = formant frequencies
F3
F2
F1
m a:
Narrow-band spectrogram Wide-band spectrogram
Amplitude
Amplitude
500
500
Frequency (Hz) Frequency (Hz)
Power Spectrum
Narrow-band spectrogram Wide-band spectrogram
F3
F3
F2
F2
F1 H2
H1
F1
m a:
Waveform of [ma:]1
1 = fundamental frequency (f0)
1 period
p
F4
F3
F2
F1
F1
F3
F2
F4
www.clas.u.edu/users/ratree/Lin_6932/Week%203%20Tape%20recording%20techniques.ppt
Measurement and Recording
Sampling rate = how oIen measure
amplitude of signal
If sampling rate is too slow, will distort or
miss high frequency signals
Get alias or false paLerns of repea.ng
cycles
Aliasing
Time
Sampling Rates for Speech
To avoid aliasing, need to sample at 2x highest
frequency you want
Then, lter out extra frequencies
For example, to accurately record frequencies
up to 5000 Hz: Sample at 10 K
Then lter out all frequencies above 5000
Nyquist or An.-aliasing Filter
Sample at 10 K, lter at 5K
Sample at 8 K, lter at 4 K
Sample at 16 K, lter at 8 K
Sample at 20 K, lter at 10 K
Sampling rate measured in sample points/
second.
We say at 5 K or 5 Khz
Quan.za.on
Concerns how many levels of amplitude
Divide dynamic range into levels
Need to have enough levels
Few levels = stair step shape--not like orig.
Few levels = quan.za.on noise
Number of bits = 2 Number of bits levels of
quan.za.on
Sampling showing Quan.zing
Steps
Rules for Digi.zing Speech
Recording device
microphone Computer
Computer Display
Speakers
Concepts for Speech/Hearing
Dynamic range = range of amplitudes or
frequencies device can handle
Dynamic range of 50 db good enough for speech
recording
Very quiet room has 30 dB signal/noise ra.o
Eliminate noise and/or improve signal/noise ra.o
= good recordings
PreEmphasis
Human Voice Shows Spectral Tilt, or Roll-o
12 Db down per octave at gloqs
6 Db up per octave from radia.on at lips
Combina.on = 6 Db per octave spectral .lt when
recorded at lips
Makes it Hard to see Higher Formants in Speech
Analysis
Addi.onal Amplica.on of Higher Formants (=
Pre-Emphasis) is added automa.cally unless you
specify you dont want it
Whats changeable? ANALYSIS
Dura.on of overall window
for calcula.onmeans how
many cycles?
Step size for calcula.on
move to next by how many
ms?
Dura.on of window for
deciding what cons.tutes a
cycle means how big is a
cycle?
Shape of window
Making a Spectrum
Window for
cycle
Step
What Period of Time is Best for
Window?