Voice
DSP
Processing
I
Yaakov J. Stein
Chief Scientist
RAD Data Communications
Stein VoiceDSP 1.1
Voice DSP
Part 1 Speech biology and what we can learn from it
Part 2 Speech DSP (AGC, VAD, features, echo cancellation)
Part 3 Speech compression techiques
Part 4 Speech Recognition
Stein VoiceDSP 1.2
Voice DSP - Part 1a
Speech production mechanisms
 Biology of the vocal tract
 Pitch and formants
 Sonograms
 The basic LPC model
 The cepstrum
 LPC cepstrum
 Line spectral pairs
Stein VoiceDSP 1.3
Voice DSP - Part 1b
Speech perception mechanisms

Biology of the ear

Psychophysical phenomena
– Weber’s law
– Fechner’s law
– Changes
– Masking
Stein VoiceDSP 1.4
Voice DSP - Part 1c
Speech quality measurement

Subjective measurement
– MOS and its variants

Objective measurement
– PSQM, PESQ
Stein VoiceDSP 1.5
Voice DSP - Part 2a
Basic speech processing
 Simplest processing
– AGC
– Simplistic VAD
 More complex processing
– pitch tracking
– formant tracking
– U/V decision
– computing LPC and other features
Stein VoiceDSP 1.6
Voice DSP - Part 2b
Echo Cancellation
 Sources of echo (acoustic vs. line echo)
 Echo suppression and cancellation
 Adaptive noise cancellation
 The LMS algorithm
 Other adaptive algorithms
 The standard LEC
Stein VoiceDSP 1.7
Voice DSP - Part 3
Speech compression techniques
 PCM
 ADPCM
 SBC
 VQ
 ABS-CELP
 MBE
 MELP
 STC
 Waveform Interpolation
Stein VoiceDSP 1.8
Voice DSP - Part 4
Speech Recognition tasks
ASR Engine
Phonetic labeling
DTW
HMM
State-of-the-Art
Stein VoiceDSP 1.9
Voice DSP - Part 1a
Speech
production
mechanisms
Stein VoiceDSP 1.10
Speech Production Organs
Brain
Hard
Palate
Nasal
cavity
Velum
Teeth
Lips
Mouth
cavity
Uvula
Pharynx
Tongue
Esophagus
Larynx
Trachea
Lungs
Stein VoiceDSP 1.11
Speech Production Organs - cont.

Air from lungs is exhaled into trachea (windpipe)

Vocal chords (folds) in larynx can produce periodic pulses of air
by opening and closing (glottis)

Throat (pharynx), mouth, tongue and nasal cavity modify air flow

Teeth and lips can introduce turbulence

Epiglottis separates esophagus (food pipe) from trachea
Stein VoiceDSP 1.12
Voiced vs. Unvoiced Speech





When vocal cords are held open air flows unimpeded
When laryngeal muscles stretch them glottal flow is in bursts
When glottal flow is periodic called voiced speech
Basic interval/frequency called the pitch
Pitch period usually between 2.5 and 20 milliseconds
Pitch frequency between 50 and 400 Hz
You can feel the vibration of the larynx


Vowels are always voiced (unless whispered)
Consonants come in voiced/unvoiced pairs
for example : B/P K/G D/T V/F J/CH TH/th W/WH Z/S ZH/SH
Stein VoiceDSP 1.13
Excitation spectra

Voiced speech
Pulse train is not sinusoidal - harmonic rich
f

Unvoiced speech
Common assumption : white noise
f
Stein VoiceDSP 1.14
Effect of vocal tract

Mouth and nasal cavities have resonances

Resonant frequencies
depend on geometry
Stein VoiceDSP 1.15
Effect of vocal tract - cont.

Sound energy at these resonant frequencies is amplified
Frequencies of peak amplification are called formants
F1
frequency response

F2
F3
F4
frequency
voiced speech
unvoiced speech
F0
Stein VoiceDSP 1.16
Formant frequencies

Peterson - Barney data (note the “vowel triangle”)
Stein VoiceDSP 1.17
Sonograms
Stein VoiceDSP 1.18
Cylinder model(s)
Rough model of throat and mouth cavity
Voice
Excitation
With nasal cavity
Voice
Excitation
open
open
open/closed
Stein VoiceDSP 1.19
Phonemes



The smallest acoustic unit that can change meaning
Different languages have different phoneme sets
Types:
(notations: phonetic, CVC, ARPABET)
– Vowels
• front (heed, hid, head, hat)
• mid (hot, heard, hut, thought)
• back (boot, book, boat)
• dipthongs (buy, boy, down, date)
– Semivowels
• liquids (w, l)
• glides (r, y)
Stein VoiceDSP 1.20
Phonemes - cont.
– Consonants
• nasals (murmurs) (n, m, ng)
• stops (plosives)
– voiced (b,d,g)
– unvoiced (p, t, k)
• fricatives
– voiced (v, that, z, zh)
– unvoiced (f, think, s, sh)
• affricatives (j, ch)
• whispers (h, what)
• gutturals ( ‫ח‬
,‫) ע‬
• clicks, etc. etc. etc.
Stein VoiceDSP 1.21
Basic LPC Model
Pulse
Generator
U/V
Switch
LPC
synthesis
filter
White Noise
Generator
Stein VoiceDSP 1.22
Basic LPC Model - cont.

Pulse generator produces a harmonic rich periodic impulse
train (with pitch period and gain)

White noise generator produces a random signal
(with gain)

U/V switch chooses between voiced and unvoiced speech

LPC filter amplifies formant frequencies
(all-pole or AR IIR filter)

The output will resemble true speech to within residual error
Stein VoiceDSP 1.23
Cepstrum
Another way of thinking about the LPC model
Speech spectrum is the obtained from multiplication
Spectrum of (pitch) pulse train times
Vocal tract (formant) frequency response
So log of this spectrum is obtained from addition
Log spectrum of pitch train plus
Log of vocal tract frequency response
Consider this log spectrum to be the spectrum of some new signal
called the cepstrum
The cepstrum is the sum of two components:
excitation plus vocal tract
Stein VoiceDSP 1.24
Cepstrum - cont.
Cepstral processing has its own language
 Cepstrum (note that this is really a signal in the time domain)

Quefrency (its units are seconds)

Liftering (filtering)

Alanysis

Saphe
Several variants:
 complex cepstrum
 power cesptrum
 LPC cepstrum
Stein VoiceDSP 1.25
Do we know enough?
Standard speech model (LPC)
(used by most speech processing/compression/recognition systems)
is a model of speech production
Unfortunately, speech production and speech perception systems
are not matched
So next we’ll look at the biology of the hearing (auditory) system
and some psychophysics (perception)
Stein VoiceDSP 1.26
Voice DSP - Part 1b
Speech
Hearing &perception
mechanisms
Stein VoiceDSP 1.27
Hearing Organs
Stein VoiceDSP 1.28
Hearing Organs - cont.











Sound waves impinge on outer ear enter auditory canal
Amplified waves cause eardrum to vibrate
Eardrum separates outer ear from middle ear
The Eustachian tube equalizes air pressure of middle ear
Ossicles (hammer, anvil, stirrup) amplify vibrations
Oval window separates middle ear from inner ear
Stirrup excites oval window which excites liquid in the cochlea
The cochlea is curled up like a snail
The basilar membrane runs along middle of cochlea
The organ of Corti transduces vibrations to electric pulses
Pulses are carried by the auditory nerve to the brain
Stein VoiceDSP 1.29
Function of Cochlea







Cochlea has 2 1/2 to 3 turns
were it straightened out it would be 3 cm in length
The basilar membrane runs down the center of the cochlea
as does the organ of Corti
15,000 cilia (hairs) contact the vibrating basilar membrane
and release neurotransmitter stimulating 30,000 auditory neurons
Cochlea is wide (1/2 cm) near oval window and tapers towards apex
is stiff near oval window and flexible near apex
Hence high frequencies cause section near oval window to vibrate
low frequencies cause section near apex to vibrate
Overlapping bank of filter frequency decomposition
Stein VoiceDSP 1.30
Psychophysics - Weber’s law
Ernst Weber Professor of physiology at Leipzig in the early 1800s
Just Noticeable Difference :
minimal stimulus change that can be detected by senses
Discovery:
DI=KI
Example
Tactile sense: place coins in each hand
subject could discriminate between with 10 coins and 11,
but not 20/21, but could 20/22!
Similarly vision lengths of lines, taste saltiness, sound frequency
Stein VoiceDSP 1.31
Weber’s law - cont.
This makes a lot of sense
Bill Gates
Stein VoiceDSP 1.32
Psychophysics - Fechner’s law
Weber’s law is not a true psychophysical law
it relates stimulus threshold to stimulus (both physical entities)
not internal representation (feelings) to physical entity
Gustav Theodor Fechner
student of Weber medicine, physics philosophy
Simplest assumption: JND is single internal unit
Using Weber’s law we find:
Y = A log I + B
Fechner Day (October 22 1850)
Stein VoiceDSP 1.33
Fechner’s law - cont.
Log is very compressive
Fechner’s law explains the fantastic ranges of our senses
Sight: single photon - direct sunlight 1015
Hearing: eardrum move 1 H atom - jet plane 1012
Bel defined to be log10 of power ratio
decibel (dB) one tenth of a Bel
d(dB) = 10 log10 P 1 / P 2
Stein VoiceDSP 1.34
Fechner’s law - sound amplitudes
Companding
adaptation of logarithm to positive/negative signals
m-law
and
A-law
are piecewise linear approximations
Equivalent to linear sampling at 12-14 bits
(8 bit linear sampling is significantly more noisy)
Stein VoiceDSP 1.35
Fechner’s law - sound frequencies
octaves, well tempered scale
12
2
Critical bands
Frequency warping
Melody
1 KHz = 1000, JND afterwards
f
M ~ 1000 log2 ( 1 + fKHz )
Barkhausen can be simultaneously heard
B ~ 25 + 75 ( 1 + 1.4 f2KHz )0.69
excite different basilar membrane regions
Stein VoiceDSP 1.36
Psychophysics - changes
Our senses respond to changes
Inverse
E
Filter
Stein VoiceDSP 1.37
Psychophysics - masking
Masking: strong tones block weaker ones at nearby frequencies
narrowband noise blocks tones (up to critical band)
f
Stein VoiceDSP 1.38
Voice DSP - Part 1c
Speech
Quality
Measurement
Stein VoiceDSP 1.39
Why does it sound
the way it sounds?
PSTN




BW=0.2-3.8 KHz, SNR>30 dB
PCM, ADPCM (BER 10-3)
five nines reliability
line echo cancellation
Voice over packet network




speech compression
delay, delay variation, jitter
packet loss/corruption/priority
echo cancellation
Stein VoiceDSP 1.40
Subjective Voice Quality
Old Measures



meet neat seat feet Pete beat heat
5/9
DRT
DAM
The modern scale


MOS
DMOS
Stein VoiceDSP 1.41
MOS according to ITU
P.800 Subjective Determination of Transmission Quality
Annex B:
Absolute Category Rating (ACR)
Listening Quality
5
4
3
2
1
excellent
good
fair
poor
bad
Listening Effort
relaxed
attention needed
moderate effort
considerable effort
no meaning
with feasible effort
Stein VoiceDSP 1.42
MOS according to ITU (cont)
Annex D Degradation Category Rating (DCR)
Annex E Comparison Category Rating (CCR)

ACR not good at high quality speech
DCR
5
4
3
2
1
0
-1
-2
-3
inaudible
not annoying
slightly annoying
annoying
very annoying
CCR
much better
better
slightly better
the same
slightly worse
worse
much worse
Stein VoiceDSP 1.43
Some MOS numbers
Effect of Speech Compression:
(from ITU-T Study Group 15)







Quiet room 48 KHz 16 bit linear sampling
PCM (A-law/mlaw) 64 Kb/s
G.723.1 @ 6.3 Kb/s
G.729 @ 8 Kb/s
5.0
4.1
3.9
3.9
ADPCM G.726 32 Kb/s
GSM @ 13Kb/s
VSELP IS54 @ 8Kb/s
3.8
3.6
3.4
toll quality
Stein VoiceDSP 1.44
The Problem(s) with MOS
Accurate MOS tests are the only reliable benchmark
BUT

MOS tests are off-line

MOS tests are slow
MOS tests are expensive
Different labs give consistently different results
Most MOS tests only check one aspect of system



Stein VoiceDSP 1.45
The Problem(s) with SNR
Naive question: Isn’t CCR the same as SNR?
SNR does not correlate well with subjective criteria
Squared difference is not an accurate comparator




Gain
Delay
Phase
Nonlinear processing
Stein VoiceDSP 1.46
Speech distance measures
Many objective measures have been proposed:





Segmental SNR
Itakura Saito distance
Euclidean distance in Cepstrum space
Bark spectral distortion
Coherence Function
None correlate well with MOS
ITU target - find a quality-measure that does correlate well
Stein VoiceDSP 1.47
Some objective methods
Perceptual Speech Quality Measurement (PSQM)
ITU-T P.861
Perceptual Analysis Measurement System (PAMS)
BT proprietary technique
Perceptual Evaluation of Speech Quality (PESQ)
ITU-T P.862
Objective Measurement of Perceived Audio Quality (PAQM)
ITU-R BS.1387
Stein VoiceDSP 1.48
Objective Quality Strategy
channel
speech
QM
QM
to
MOS
MOS
estimate
Stein VoiceDSP 1.49
PSQM philosophy
(from P.861)
Internal
Perceptual
Representation
model
Audible
Cognitive
Difference
Model
Perceptual
model
Internal
Representation
Stein VoiceDSP 1.50
PSQM philosophy (cont)
Perceptual Modelling



(Internal representation)
Short time Fourier transform
Frequency warping (telephone-band filtering, Hoth noise)
Intensity warping
Cognitive Modelling




Loudness scaling
Internal cognitive noise
Asymmetry
Silent interval processing
PSQM Values

0 (no degradation) to 6.5 (maximum degradation)
Conversion to MOS


PSQM to MOS calibration using known references
Equivalent Q values
Stein VoiceDSP 1.51
Problems with PSQM
Designed for telephony grade speech codecs
Doesn’t take network effects into account:



filtering
variable time delay
localized distortions
Draft standard P.862 adds:



transfer function equalization
time alignment, delay skipping
distortion averaging
Stein VoiceDSP 1.52
PESQ philosophy
(from P.862)
Perceptual
Internal
model
Representation
Time
Audible
Cognitive
Alignment
Difference
Model
Perceptual
Internal
model
Representation
Stein VoiceDSP 1.53
Descargar

Voice DSP Processing