224s.09.lec10-1

# 224s.09.lec10-1 - CS224S/LINGUIST281...

This preview shows pages 1–12. Sign up to view the full content.

CS 224S / LINGUIST 281 Speech Recognition, Synthesis, and  Dialogue Dan Jurafsky Lecture 10: Acoustic Modeling IP No tic e :

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Outline for Today Speech Recognition Architectural Overview Hidden Markov Models in general and for  speech Forward Viterbi Decoding How this fits into the ASR component of course Jan 27 HMMs, Forward, Viterbi, Jan 29 Baum-Welch (Forward-Backward) Feb 3: Feature Extraction, MFCCs, start of AM (VQ) Feb 5: Acoustic Modeling: GMMs Feb 10: N-grams and Language Modeling Feb 24: Search and Advanced Decoding Feb 26: Dealing with Variation
Outline for Today Acoustic Model Increasingly sophisticated models Acoustic Likelihood for each state: Gaussians Multivariate Gaussians Mixtures of Multivariate Gaussians Where a state is progressively: CI Subphone (3ish per phone) CD phone (=triphones) State-tying of CD phone If Time: Evaluation Word Error Rate

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Reminder: VQ To compute p(o t |q j ) Compute distance between feature vector o t   and each codeword (prototype vector) in a preclustered codebook where distance is either Euclidean Mahalanobis Choose the vector that is the closest to o t and take its codeword v k And then look up the likelihood of v k  given HMM state  j in the B matrix B j (o t )=b j (v k ) s.t. v k  is codeword of closest vector  to o t Using Baum-Welch as above
Computing b j (v k ) feature value 1 for state j feature value 2 for state j • b j (v k ) = number of vectors with codebook index k in state j number of vectors in state j = = 14 1 56 4 Slide from John-Paul Hosum, OHSU/OGI

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Summary: VQ Training: Do VQ and then use Baum-Welch to assign  probabilities to each symbol Decoding: Do VQ and then use the symbol probabilities  in decoding
Directly Modeling Continuous  Observations Gaussians Univariate Gaussians Baum-Welch for univariate Gaussians Multivariate Gaussians Baum-Welch for multivariate Gausians Gaussian Mixture Models (GMMs) Baum-Welch for GMMs

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Better than VQ VQ is insufficient for real ASR Instead: Assume the possible values of the  observation feature vector o t  are normally  distributed. Represent the observation likelihood function  b j (o t ) as a Gaussian with mean  μ j  and variance  σ j 2 f ( x | μ , σ ) = 1 2 π exp( - ( x ) 2 2 2 )
Gaussians are parameters by mean  and variance

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
For a discrete random variable X Mean is the expected value of X  Weighted sum over the values of X Variance is the squared average deviation  from mean Reminder: means and  variances
Gaussian as Probability Density  Function

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
This is the end of the preview. Sign up to access the rest of the document.

## This note was uploaded on 04/21/2011 for the course CS 224 taught by Professor De during the Spring '11 term at Kentucky.

### Page1 / 59

224s.09.lec10-1 - CS224S/LINGUIST281...

This preview shows document pages 1 - 12. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online