Sound, Mixtures, and Learning
Dan Ellis
<dpwe@ee.columbia.edu>
Laboratory for Recognition and Organization of Speech and Audio
(LabROSA)
Electrical Engineering, Columbia University
http://labrosa.ee.columbia.edu/
Outline
1
Auditory Scene Analysis
2
Speech Recognition & Mixtures
3
Fragment Recognition
4
Alarm Sound Detection
5
Future Work
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 1/37
Auditory Scene Analysis
1
4000
frq/Hz
3000
0
2000
-20
1000
-40
0
0
2
4
6
8
10
12
time/s
-60
level / dB
Analysis
Voice (evil)
Voice (pleasant)
Stab
Rumble
Choir
Strings
•
Auditory Scene Analysis: describing a complex
sound in terms of high-level sources/events
- ... like listeners do
•
Hearing is ecologically grounded
- reflects ‘natural scene’ properties
- subjective, not absolute
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 2/37
Sound, mixtures, and learning
freq / kHz
Speech
Noise
Speech + Noise
4
3
+
2
=
1
0
0
0.5
1
time / s
0
0.5
1
time / s
0
0.5
1
•
Sound
- carries useful information about the world
- complements vision
•
Mixtures
- .. are the rule, not the exception
- medium is ‘transparent’, sources are many
- must be handled!
•
Learning
- the ‘speech recognition’ lesson:
let the data do the work
- like listeners
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 3/37
time / s
The problem with recognizing mixtures
“Imagine two narrow channels dug up from the edge of a
lake, with handkerchiefs stretched across each one.
Looking only at the motion of the handkerchiefs, you are
to answer questions such as: How many boats are there
on the lake and where are they?” (after Bregman’90)
•
Received waveform is a mixture
- two sensors, N signals ... underconstrained
•
Disentangling mixtures as the primary goal?
- perfect solution is not possible
- need experience-based constraints
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 4/37
Human Auditory Scene Analysis
(Bregman 1990)
•
How do people analyze sound mixtures?
- break mixture into small elements (in time-freq)
- elements are grouped in to sources using cues
- sources have aggregate attributes
•
Grouping ‘rules’ (Darwin, Carlyon, ...):
- cues: common onset/offset/modulation,
harmonicity, spatial location, ...
Onset
map
Frequency
analysis
Harmonicity
map
Source
properties
Grouping
mechanism
Position
map
(after Darwin, 1996)
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 5/37
Cues to simultaneous grouping
freq / Hz
•
Elements + attributes
8000
6000
4000
2000
0
0
1
2
3
4
5
6
7
8
9 time / s
•
Common onset
- simultaneous energy has common source
•
Periodicity
- energy in different bands with same cycle
•
Other cues
- spatial (ITD/IID), familiarity, ...
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 6/37
The effect of context
Context can create an ‘expectation’:
i.e. a bias towards a particular interpretation
•
e.g. Bregman’s “old-plus-new” principle:
A change in a signal will be interpreted as an
added source whenever possible
frequency / kHz
•
2
1
+
0
0.0
0.4
0.8
1.2
time / s
- a different division of the same energy
depending on what preceded it
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 7/37
Computational Auditory Scene Analysis
(CASA)
CASA
Object 1 description
Object 2 description
Object 3 description
...
•
Goal: Automatic sound organization ;
Systems to ‘pick out’ sounds in a mixture
- ... like people do
•
E.g. voice against a noisy background
- to improve speech recognition
•
Approach:
- psychoacoustics describes grouping ‘rules’
- ... just implement them?
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 8/37
The Representational Approach
(Brown & Cooke 1993)
•
input
mixture
Implement psychoacoustic theory
signal
features
Front end
(maps)
Object
formation
discrete
objects
Source
groups
Grouping
rules
freq
onset
time
period
frq.mod
- ‘bottom-up’ processing
- uses common onset & periodicity cues
•
frq/Hz
Able to extract voiced speech:
brn1h.aif
frq/Hz
3000
3000
2000
1500
2000
1500
1000
1000
600
600
400
300
400
300
200
150
200
150
100
brn1h.fi.aif
100
0.2
0.4
0.6
0.8
1.0
time/s
Sound, mixtures, learning @ OSU - Dan Ellis
0.2
0.4
0.6
0.8
2002-08-10 - 9/37
1.0
time/s
Restoration in sound perception
•
Auditory ‘illusions’ = hearing what’s not there
•
The continuity illusion
f/Hz
ptshort
4000
2000
1000
0.0
•
0.2
0.4
0.6
0.8
1.0
1.2
1.4
time/s
SWS
f/Bark
15
80
60
S1−env.pf:0
10
5
40
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
- duplex perception
•
How to model in CASA?
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 10/37
Adding top-down constraints
Perception is not direct
but a search for plausible hypotheses
•
Data-driven (bottom-up)...
input
mixture
Front end
signal
features
Object
formation
discrete
objects
Grouping
rules
Source
groups
- objects irresistibly appear
vs. Prediction-driven (top-down)
hypotheses
Noise
components
Hypothesis
management
prediction
errors
input
mixture
Front end
signal
features
Compare
& reconcile
Predict
& combine
Periodic
components
predicted
features
- match observations
with parameters of a world-model
- need world-model constraints...
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 11/37
Approaches to sound mixture recognition
•
Recognize combined signal
- ‘multicondition training’
- combinatorics..
•
Separate signals
- e.g. CASA, ICA
- nice, if you can do it
•
Segregate features into fragments
- then missing-data recognition
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 12/37
Aside: Evaluation
•
Evaluation is a big problem for CASA
- what is the goal, really?
- what is a good test domain?
- how do you measure performance?
•
SNR improvement
- not easy given only before-after signals:
correspondence problem
- can do with fixed filtering mask;
rewards removing signal as well as noise
•
ASR improvement
- recognizers typically very sensitive to artefacts
•
‘Real’ task?
- mixture corpus with specific sound events...
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 13/37
Outline
1
Auditory Scene Analysis
2
Speech Recognition & Mixtures
- standard ASR
- approaches to speech + noise
3
Fragment Recognition
4
Alarm Sound Detection
5
Future Work
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 14/37
Speech recognition & mixtures
2
•
Speech recognizers are the most successful
and sophisticated acoustic recognizers to date
sound
Feature
calculation
D AT A
feature vectors
Acoustic model
parameters
Word models
s
ah
t
Language model
p("sat"|"the","cat")
p("saw"|"the","cat")
•
Acoustic
classifier
phone probabilities
HMM
decoder
phone / word
sequence
Understanding/
application...
‘State of the art’ word-error rates (WERs):
- 2% (dictation) - 30% (phone conv’ns)
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 15/37
Learning acoustic models
•
Goal: describe p ( X M ) with e.g. GMMs
p(x|ω1)
Labeled
training
examples
{xn,ωxn}
Sort
according
to class
Estimate
conditional pdf
for class ω1
•
Separate models for each class
- generalization as blurring
•
Training data labels from:
- manual annotation
- ‘best path’ from earlier classifier (Viterbi)
- EM: joint estimation of labels & pdfs
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 16/37
Speech + noise mixture recognition
•
Background noise is biggest (?) problem
facing current ASR
•
Feature invariance approach:
Design features to reflect only speech
- e.g. normalization, mean subtraction
•
Ideally, models of clean speech will match
speech in noise
- .. although training on noisy examples can’t hurt
•
Static noise is relatively easy
- but: non-static noise?
•
Alternative:
More complex models of the signal
- separate models for speech and ‘rest’
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 17/37
HMM decomposition
(e.g. Varga & Moore 1991, Roweis 2000)
•
Total signal model has independent state
sequences for 2+ component sources
model 2
model 1
observations / time
•
New combined state space q' = {q1 q2}
- new observation pdfs for each combination
p ( X q 1, q 2 )
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 18/37
Problems with HMM decomposition
•
O(qk)N is exponentially large...
•
Feature normalization no longer holds!
- each source has a different gain
→ model at various SNRs?
- models typically don’t use overall energy C0
- each source has a different channel H[k]
•
Modeling every possible sub-state combination
is inefficient, inelegant and impractical
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 19/37
Outline
1
Auditory Scene Analysis
2
Speech Recognition & Mixtures
3
Fragment Recognition
- separating signals vs. separating features
- missing data recognition
- recognizing multiple sources
4
Alarm Sound Detection
5
Future Work
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 20/37
Fragment Recognition
3
(Jon Barker & Martin Cooke, Sheffield)
•
Signal separation is too hard!
Instead:
- segregate features into partially-observed
sources
- then classify
•
Made possible by ‘missing data’ recognition
- integrate over uncertainty in observations
for optimal posterior distribution
•
Goal:
Relating clean speech models P(X|M)
to speech + noise mixture observations
- .. and making it tractable
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 21/37
Comparing different segregations
•
Standard classification chooses between
models M to match source features X
P( M )
M ∗ = argmax P ( M X ) = argmax P ( X M ) ⋅ -------------P( X )
M
M
•
Mixtures → observed features Y, segregation S,
all related by P ( X Y , S )
Observation
Y(f )
Source
X(f )
Segregation S
freq
- spectral features allow clean relationship
•
Joint classification of model and segregation:
P( X Y , S )
P ( M , S Y ) = P ( M ) P ( X M ) ⋅ ------------------------- dX ⋅ P ( S Y )
P( X )
∫
- integral collapses in several cases...
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 22/37
Calculating fragment matches
P( X Y , S )
P ( M , S Y ) = P ( M ) P ( X M ) ⋅ ------------------------- dX ⋅ P ( S Y )
P( X )
∫
•
P(X|M) - the clean-signal feature model
•
P(X|Y,S)/P(X) - is X ‘visible’ given segregation?
•
Integration collapses some channels...
•
P(S|Y) - segregation inferred from observation
- just assume uniform, find S for most likely M
- use extra information in Y to distinguish S’s
e.g. harmonicity, onset grouping
•
Result:
- probabilistically-correct relation between
clean-source models P(X|M)
and inferred contributory source P(M,S|Y)
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 23/37
Speech fragment decoder results
•
Simple P(S|Y) model forces contiguous regions
to stay together
- big efficiency gain when searching S space
"1754" + noise
90
AURORA 2000 - Test Set A
80
WER / %
70
SNR mask
60
HTK clean training
MD Soft SNR
HTK multicondition
50
40
30
20
Fragments
10
0
-5
Fragment
Decoder
•
0
5
10
SNR / dB
15
"1754"
Clean-models-based recognition
rivals trained-in-noise recognition
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 24/37
20
clean
Multi-source decoding
•
Search for more than one source
q2(t)
S2(t)
Y(t)
S1(t)
q1(t)
•
Mutually-dependent data masks
•
Use e.g. CASA features to propose masks
- locally coherent regions
•
Theoretical vs. practical limits
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 25/37
Outline
1
Auditory Scene Analysis
2
Speech Recognition & Mixtures
3
Fragment Recognition
4
Alarm Sound Detection
- sound
- mixtures
- learning
5
Future Work
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 26/37
Alarm sound detection
4
•
freq / kHz
s0n6a8+20
4
Alarm sounds have particular structure
- people ‘know them when they hear them’
- clear even at low SNRs
hrn01
bfr02
buz01
20
3
0
2
-20
1
0
0
5
10
15
20
25
time / s
•
Why investigate alarm sounds?
- they’re supposed to be easy
- potential applications...
•
Contrast two systems:
- standard, global features, P(X|M)
- sinusoidal model, fragments, P(M,S|Y)
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 27/37
-40
level / dB
freq / Hz
Alarms: Sound (representation)
•
Standard system: Mel Cepstra
- have to model alarms in noise context:
each cepstral element depends on whole signal
•
Contrast system: Sinusoid groups
- exploit sparse, stable nature of alarm sounds
- 2D-filter spectrogram to enhance harmonics
- simple magnitude threshold, track growing
- form groups based on common onset
5000
5000
5000
4000
4000
4000
3000
3000
3000
2000
2000
2000
1000
1000
1000
0
•
1
1.5
2
2.5
0
1
1.5
2
time / sec
2.5
0
1
1
1.5
2
2.5
Sinusoid representation is already fragmentary
- does not record non-peak energies
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 28/37
Alarms: Mixtures
•
Effect of varying SNR on representations:
- sinusoid peaks have ~ invariant properties
Sine track groups
Cepstra (normalized)
10 dB SNR 60 dB SNR
2
10
1
5
0
2
3
10
1
5
0
1
3
0 dB SNR
2
5
4
0
0
5
10
15
20
25
time / s
Sound, mixtures, learning @ OSU - Dan Ellis
0
5
10
15
20
2002-08-10 - 29/37
25
-5
time / s
Alarms: Learning
•
Standard: train MLP on noisy examples
Feature
extraction
Sound
mixture
•
PLP
cepstra
Neural net
acoustic
classifier
Detected
alarms
Alarm14
non alarm
12
Magnitude SD
80
Spectral moment
Alarm
probability
Alternate: learn distributions of group features
- duration, frequency deviation, amp. modulation...
100
60
40
20
0
Median
filtering
10
8
6
4
2
0
50
100
150
Spectral centroid
200
0
0
10
20
30
Inverse Frequency SD
- underlying models are clean (isolated)
- recognize in different contexts...
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 30/37
40
freq / kHz
Alarms: Results
Restaurant+ alarms (snr 0 ns 6 al 8)
4
3
2
1
0
MLP classifier output
freq / kHz
0
Sound object classifier output
4
6
9
8
7
3
2
1
0
20
•
25
30
35
40
45
time/sec 50
Both systems commit many insertions at 0dB
SNR, but in different circumstances:
Noise
Neural net system
Del
Ins
Tot
Sinusoid model system
Del
Ins
Tot
1 (amb)
7 / 25
2
36%
14 / 25
1
60%
2 (bab)
5 / 25
63
272%
15 / 25
2
68%
3 (spe)
2 / 25
68
280%
12 / 25
9
84%
4 (mus)
8 / 25
37
180%
9 / 25
135
576%
170
192%
50 / 100
147
197%
Overall 22 / 100
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 31/37
Alarms: Summary
•
Sinusoid domain
- feature components belong to 1 source
- simple ‘segregation’ (grouping) model
- alarm model as properties of group
- robust to partial feature observation
•
Future improvements
- more complex alarm class models
- exploit repetitive structure of alarms
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 32/37
Outline
1
Auditory Scene Analysis
2
Speech Recognition & Mixtures
3
Fragment Recognition
4
Alarm Sound Detection
5
Future Work
- generative models & inference
- model acquisition
- ambulatory audio
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 33/37
Future work
5
•
CASA as generative model parameterization:
Generation
model
Channel
parameters
Source
models
Source
signals
M1
Y1
Received
signals
X1
C2
Model
dependence
Θ
C1
Observations
M2
Y2
X2
O
Analysis
structure
Observations
O
Fragment
formation
Mask
allocation
p(X|Mi,Ki)
Sound, mixtures, learning @ OSU - Dan Ellis
{Ki}
Model
fitting
{Mi}
Likelihood
evaluation
2002-08-10 - 34/37
Learning source models
•
The speech recognition lesson:
Use the data as much as possible
- what can we do with unlimited data feeds?
•
Data sources
- clean data corpora
- identify near-clean segments in real sound
- build up ‘clean’ views from partial observations?
•
Model types
- templates
- parametric/constraint models
- HMMs
•
Hierarchic classification
vs. individual characterization...
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 35/37
Personal Audio Applications
•
Smart PDA records everything
•
Only useful if we have index, summaries
- monitor for particular sounds
- real-time description
•
Scenarios
- personal listener → summary of your day
- future prosthetic hearing device
- autonomous robots
•
Meeting data, ambulatory audio
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 36/37
Summary
•
Sound
- carries important information
•
Mixtures
- need to segregate different source properties
- fragment-based recognition
•
Learning
- information extracted by classification
- models guide segregation
•
Alarm sounds
- simple example of fragment recognition
•
General sounds
- recognize simultaneous components
- acquire classes from training data
- build index, summary of real-world sound
Sound, mixtures, learning @ OSU - Dan Ellis
2002-08-10 - 37/37