automatic speech recognition using an echo state network mark d. skowronski computational...

Automatic speech recognition using an echo state network

Mark D. SkowronskiComputational Neuro-Engineering LabElectrical and Computer Engineering

University of Florida, Gainesville, FL, USAMay 10, 2006

CNEL Seminar History• Ratio spectrum, Oct. 2000• HFCC, Sept. 2002• Bats, Dec. 2004• Electrohysterography, Aug. 2005• Echo state network, May 2006

2006

2000

Overview• ASR motivations• Intro to echo state network• Multiple readout filters• ASR experiments• Conclusions

ASR Motivations• Speech is most natural form of communication

among humans.• Human-machine interaction lags behind with

tactile interface.• Bottleneck in machine understanding is signal-

to-symbol translation.• Human speech a “tough” signal:

– Nonstationary– Non-Gaussian– Nonlinear systems for production/perception

How to handle the “non”-ness of speech?

ASR State of the Art• Feature extraction: HFCC

– bio-inspired frequency analysis– tailored for statistical models

• Acoustic pattern rec: HMM– Piecewise-stationary stochastic model– Efficient training/testing algorithms– …but several simplistic assumptions

• Language models– Uses knowledge of language, grammar– HMM implementations– Machine language understanding still elusive (spam

blockers)

Hidden Markov ModelPremier stochastic model of non-stationary time series used for decision making.

Assumptions:1) Speech is piecewise-stationary process.2) Features are independent.3) State duration is exponential.4) State transition prob. function of previous-next state only.

Can we devise a better pattern recognition model?

Echo State Network

• Partially trained recurrent neural network, Herbert Jaeger, 2001

• Unique characteristics:– Recurrent “reservoir” of processing elements,

interconnected with random untrained weights.– Linear readout weights trained with simple

regression provide closed-form, stable, unique solution.

ESN Diagram & Equations

)()())()1(()(

nnnnfn

out

in

xWyuWxWx

ESN Matrices

• Win: untrained, M x Min matrix– Zero mean, unity variance normally distributed– Scaled by rin

• W: untrained, M x M matrix– Zero mean, unity variance normally distributed– Scaled such that spectral radius r < 1

• Wout: trained, linear regression, Mout x M matrix– Regression closed-form, stable, unique solution– O(M2) per data point complexity

Echo States Conditions• The network has echo states if x(n) is

uniquely determined by left-infinite input sequence …,u(n-1),u(n).

• x(n) is an “echo” of all previous inputs.• If f is tanh activation function:

– If σmax(W)=||W||<1, guarantees echo states

– If r=|λmax(W)|>1, guarantees no echo states

ESN Training• Minimize mean-squared error between y(n)

and desired signal d(n).

))()(())()(( 1

1

TTout

out

nnnn dxxxW

pRW

Wiener solution:

ESN Example: Mackey-GlassM=60 PEsr=0.9rin=0.3u(n): MG,10000 samplesd(n)=u(n+1)

Prediction Gain(var(u)/var(e)):Input: 16.3 dBWiener: 45.1 dBESN: 62.6 dB

Multiple Readout Filters• Wout projects reservoir space to output space.• Question: how to divide reservoir space and

use multiple readout filters?• Answer: competitive network of filters

• Question: how to train/test competitive network of K filters?

• Answer: mimic HMM.

],1[),()( Kknn kout

k xWy

HMM vs. ESN ClassifierHMM ESN Classifier

Output Likelihood MSE

Architecture States, left-to-right States, left-to-right

Minimum element

Gaussian kernel Readout filter

Elements combined

GMM Winner-take-all

Transitions State transition matrix Binary switching matrix

Training Segmental K-means (Baum-Welch)

Segmental K-means

Discriminatory No Maybe, depends on desired signal

Segmental K-means: InitFor each input, xi(n) and desired di(n) for sequence i:

Divide x,d into equal-sized chunks Xη,Dη (one per state).For each n, select k(n)[1,K] uniform random.

After init. with all sequences:

Tii

nk

Tii

nk

nn

nn

))(()(

))(()(),(

),(

DXB

XXA

,1,, )( kkkout BAW

Segmental K-means: Training• For each utterance:

– Produce MSE for each readout filter.– Find Viterbi path through MSE matrix.– Use features from each state to update

auto- and cross-correlation matrices.• After all utterances: Wiener solution• Guaranteed to converge to local

minimum in MSE over training set.

ASR Example 1• Isolated English digits “zero” - “nine” from TI46 corpus: 8

male, 8 female, 26 utterances each, 12.5 kHz sampling rate.• ESN: M=60 PEs, r=2.0, rin=0.1, 10 word models, various

#states and #filters per state.• Features: 13 HFCC, 100 fps, Hamming window, pre-emphasis

(α=0.95), CMS, Δ+ΔΔ (±4 frames)• Pre-processing: zero-mean and whitening transform• M1/F1: testing; M2/F2: validation; M3-M8/F3-F8 training• Two to six training epochs for all models• Desired: next frame of 39-dimension features• Test: corrupted by additive noise from “real” sources (subway,

babble, car, exhibition hall, restaurant, street, airport terminal, train station)

• Baseline: HMM with identical input features

ASR Results, noise freeK

ESN (HMM) 1 2 3 4 5 10

Nst=1 7(171) 6(136) 3(65) 2(33) 3(4) 2(2)

2 1(83) 1(46) 0(4) 1(3) 2(2) 1(0)

3 0(126) 1(4) 0(2) 0(2) 0(1) 2(0)

5 1(11) 1(2) 0(0) 0(0) 1(0) 0(0)

10 1(2) 1(0) 1(0) 1(0) 0(0) 0(0)

15 0(1) 0(0) 0(0) 0(0) 0(0) 1(0)

20 0 0 0 0 0 1

Number of classification errors out of 518 (smaller is better).

ASR Results, noisyK

ESN (HMM) 1 2 3 4 5 10

Nst=1 70.9(22.4) 70.0(29.7) 74.6(45.6) 74.3(46.0) 74.3(36.2) 75.8(50.9)

2 76.3(41.5) 77.6(47.6) 78.3(50.1) 77.7(53.8) 77.1(50.2) 75.8(64.5)

3 78.8(29.2) 79.2(44.6) 79.3(51.7) 79.2(58.6) 79.1(58.6) 78.8(55.6)

5 81.4(51.6) 81.1(56.4) 81.6(59.7) 81.9(59.2) 81.3(59.2) 81.3(53.5)

10 84.6(57.2) 84.4(61.1) 84.4(58.7) 83.6(55.7) 83.5(56.2) 81.0(52.2)

15 85.4(64.0) 85.1(62.0) 85.0(59.2) 83.8(56.4) 82.8(52.9) 78.4(52.2)

20 85.8 85.6 84.0 83.5 82.5 72.3

Average accuracy (%),all noise sources, 0-20 dB SNR (larger is better):

ASR Results, noisySingle mixture per state (K=1): ESN classifier

ASR Results, noisySingle mixture per state (K=1): HMM baseline

ASR Example 2• Same experiment setup as Example 1.• ESN: M=600 PEs, 10 states, 1 filter per state,

rin=0.1, various r.• Desired: one-of-many encoding of class, ±1, tanh

output activation function AFTER linear readout filter.

• Test: corrupted by additive speech-shaped noise• Baseline: HMM with identical input features

ASR Results, noisy

Discussion• What gives the ESN classifier its noise-

robust characteristics?• Theory: ESN reservoir provides context

of noisy input, allowing reservoir to reduce effects of noise by averaging.

• Theory: Non-linearity and high-dimensionality of network increases linear separability of classes in reservoir space.

Future Work• Replace winner-take-all with mixture-of-

experts.• Replace segmental K-means with Baum-

Welch-type training algorithm.• “Grow” network during training.• Consider nonlinear activation functions

(e.g., tanh, softmax) AFTER linear readout filter.

Conclusions• ESN classifier using inspiration from HMM:

– Multiple readout filters per state, multiple states.– Trained as competitive network of filters.– Segmental K-means guaranteed to converge to

local minimum of total MSE from training set.• ESN classifier noise robust compared to HMM:

– Ave. over all sources, 0-20 dB SNR: +21 percentage points

– Ave. over all sources: +9 dB SNR