cs 124/linguist 180: from languages to information dan jurafsky lecture 19: speech recognition

CS 124/LINGUIST 180: From Languages to

Information

Dan JurafskyLecture 19: Speech Recognition

The final exam

• Friday March 20 12:15-3:15 in 300-300• Open book and open note• You won’t need a calculator• Computers are ok to read, e.g., the slides

and the textbooks, but no use of the internet on your laptop or any internet-aware devices, on the honor code i.e., open book and notes, but not open-web

• The problems will be very much like homework 5, which I gave you specifically to be a finals prep

Topics we covered

• http://www.stanford.edu/class/linguist180/

Some classes in these areas

cs224N Natural Language Processing Manning, Spring 2009cs224S Speech Recognition, Understanding,Dialogue, Jurafsky, cs224U Natural Language Understanding possible winter 2010ling284 History of Comp. Linguistics/NLP Jurafsky and Kay, A or

Wcs121 Intro to AI Latombe Spring 2009cs221 Artificial Intelligence Ngcs228 Structured Probabilistic ModelsKoller cs262 Computational Genomics Batzoglou Often Wintercs229 Machine Learning Ngcs270 Intro to Biomedical Informatics: Musencs276 Information Retrieval and Web Search

Manning/Raghavan

Outline for ASR

• ASR Tasks and Architecture• Five easy pieces of an ASR system

1) The Lexicon (An HMM with phones as hidden states)

2) The Language Model3) The Acoustic Model (phone detector)4) Feature extraction (“MFCC”)5) HMM Stuff:

1) Viterbi decoding2) EM (Baum-Welch) training

Applications of Speech Recognition/Understanding (ASR/ASU)

• Dictation• Telephone-based Information

GOOG 411 Directions, air travel, banking, etc

• “Google Voice” Voice mail transcription http://www.nytimes.com/2009/03/12/techn

ology/personaltech/12pogue.html

• Hands-free (in car)• Second language ('L2') (accent

reduction)• Audio archive searching and aligning1/5/07

Speaker Recognition tasks

Applications of Speaker Recognition and Language Recognition

• Language recognition for call routing• Speaker Recognition:

Speaker verification (binary decision) Voice password, telephone assistant

Speaker identification (one of N) Criminal investigation

1/5/07

Speech synthesis

• Telephone dialogue systems• Games• The new ipod shuffle

http://www.apple.com/ipodshuffle/voiceover.html

• The Kindle voice Controversy!!!

• Compare to state-of-the-art synthesis: http://www.research.att.com/~ttsweb/tts

/demo.php

LVCSR

• Large Vocabulary Continuous Speech Recognition ~20,000-64,000 words Speaker independent (vs. speaker-

dependent) Continuous speech (vs isolated-word)

• Useful for: Dictation Voice-mail transcription

Current error rates

Task Vocabulary Error Rate%

Digits 11 0.5

WSJ read speech 5K 3

WSJ read speech 20K 3

Broadcast news 64,000+ 10

Conversational Telephone

64,000+ 20

Ballpark numbers; exact numbers depend very much on the specific corpus

HSR versus ASR

• Conclusions:Machines about 5 times worse than humansGap increases with noisy speechThese numbers are rough, take with grain of

salt

Task Vocab

ASR Hum SR

Continuous digits

11 .5 .009

WSJ 1995 clean 5K 3 0.9

WSJ 1995 w/noise

5K 9 1.1

SWBD 2004 65K 20 4

Why is conversational speech harder?

• A piece of an utterance without context

• The same utterance with more context

LVCSR Design Intuition

• Build a statistical model of the speech-to-words process

• Collect lots and lots of speech, and transcribe all the words.

• Train the model on the labeled speech

• Paradigm: Supervised Machine Learning + Search

The Noisy Channel Model

• Search through space of all possible sentences.

• Pick the one that is most probable given the waveform.

The Noisy Channel Model (II)

• What is the most likely sentence out of all sentences in the language L given some acoustic input O?

• Treat acoustic input O as sequence of individual observations O = o1,o2,o3,…,ot

• Define a sentence as a sequence of words: W = w1,w2,w3,…,wn

Noisy Channel Model (III)

• Probabilistic implication: Pick the highest prob S:

• We can use Bayes rule to rewrite this:

• Since denominator is the same for each candidate sentence W, we can ignore it for the argmax:

ˆ W argmaxW L

P(W | O)

ˆ W argmaxW L

P(O |W )P(W )

ˆ W argmaxW L

P(O |W )P(W )

P(O)

Noisy channel model

ˆ W argmaxW L

P(O |W )P(W )

likelihood prior

Starting with the HMM Lexicon

• A list of words• Each one with a pronunciation in

terms of phones• We get these from on-line

pronucniation dictionary• CMU dictionary: 127K words

http://www.speech.cs.cmu.edu/cgi-bin/cmudict

• We’ll represent the lexicon as an HMM

ARPAbet

• http://www.stanford.edu/class/cs224s/arpabet.html

1/5/07

HMMs for speech: the word “six”

• Hidden states are phones:

• Loopbacks because a phone is ~100 milliseconds long An observation of speech every 10 ms So each phone repeats ~10 times

(simplifying greatly)

HMM for the digit recognition task

The rest

• So if it’s just HMMs• I just need to tell you how to build a

phone detector• The rest is the same as P-o-s or

Named Entity tagging• Phone detection algorithm:• Supervised machine learning

Classifier: Gaussian Mixture Model Features: “Mel-frequence Cepstral

Coefficients”, MFCC

Speech Production Process

• Respiration: We (normally) speak while breathing out.

Respiration provides airflow. “Pulmonic egressive airstream”

• Phonation Airstream sets vocal folds in motion. Vibration

of vocal folds produces sounds. Sound is then modulated by:

• Articulation and Resonance Shape of vocal tract, characterized by: Oral tract

Teeth, soft palate (velum), hard palate Tongue, lips, uvula

Nasal tract1/5/07

Text adopted from Sharon Rose

1/5/07

Nasal Cavity

Pharynx

Vocal Folds (within the Larynx)

Trachea

Lungs

Text copyright J. J. Ohala, Sept 2001, from Sharon Rose slide

Sagittal section of the vocal tract (Techmer 1880)

1/5/07

From Mark Liberman’s website, from Ultimate Visual Dictionary

1/5/07

From Mark Liberman’s Web Site, from Language Files (7th ed)

Vocal tract

1/5/07 Figure thnx to John Coleman!!

Vocal tract movie (high speed x-ray)

1/5/07 Figure of Ken Stevens, from Peter Ladefoged’s web site

1/5/07 Figure of Ken Stevens, labels from Peter Ladefoged’s web site

USC’s SAIL LabShri Narayanan

1/5/07

Larynx and Vocal Folds

• The Larynx (voice box) A structure made of cartilage and muscle Located above the trachea (windpipe) and

below the pharynx (throat) Contains the vocal folds (adjective for larynx: laryngeal)

• Vocal Folds (older term: vocal cords) Two bands of muscle and tissue in the

larynx Can be set in motion to produce sound

(voicing)1/5/07 Text from slides by Sharon Rose UCSD LING 111 handout

The larynx, external structure, from front

1/5/07 Figure thnx to John Coleman!!

Vertical slice through larynx, as seen from back

1/5/07 Figure thnx to John Coleman!!

Voicing:

1/5/07

•Air comes up from lungs•Forces its way through vocal cords, pushing open (2,3,4)•This causes air pressure in glottis to fall, since:

• when gas runs through constricted passage, its velocity increases (Venturi tube effect)• this increase in velocity results in a drop in pressure (Bernoulli principle)

•Because of drop in pressure, vocal cords snap together again (6-10)•Single cycle: ~1/100 of a second.

Figure & text from John Coleman’s web site

Voicelessness

• When vocal cords are open, air passes through unobstructed

• Voiceless sounds: p/t/k/s/f/sh/th/ch• If the air moves very quickly, the

turbulence causes a different kind of phonation: whisper

1/5/07

Vocal folds open during breathing

1/5/07 From Mark Liberman’s web site, from Ultimate Visual Dictionary

Vocal Fold Vibration

1/5/07

UCLA Phonetics Lab Demo

Consonants and Vowels

• Consonants: phonetically, sounds with audible noise produced by a constriction

• Vowels: phonetically, sounds with no audible noise produced by a constriction

• (it’s more complicated than this, since we have to consider syllabic function, but this will do for now)

1/5/07 Text adapted from John Coleman

Acoustic Phonetics

• Sound Waves http://www.kettering.edu/~drussell/Dem

os/waves-intro/waves-intro.html

Simple Periodic Waves (sine waves)

Time (s)0 0.02

–0.99

0.99

0

• Characterized by:• period: T• amplitude A• phase

• Fundamental frequencyin cycles per second, or Hz

• F0=1/T1 cycle

Simple periodic waves

• Computing the frequency of a wave: 5 cycles in .5 seconds = 10 cycles/second = 10 Hz

• Amplitude: 1

• Equation: Y = A sin(2ft)

Speech sound waves

• A little piece from the waveform of the vowel [iy]• Y axis:

Amplitude = amount of air pressure at that time point

Positive is compression Zero is normal air pressure, negative is rarefaction

• X axis: time.

Digitizing Speech

Digitizing Speech

• Analog-to-digital conversion• Or A-D conversion.• Two steps

Sampling Quantization

Sampling

• Measuring amplitude of signal at time t• The sampling rate needs to have at least two samples for each cycle

• Roughly speaking, one for the positive and one for the negative half of each cycle.• More than two sample per cycle is ok• Less than two samples will cause frequencies to be missed• So the maximum frequency that can be measured is one that is half the sampling rate.• The maximum frequency for a given sampling rate called Nyquist frequency

Sampling

If measure at green dots, will see a lower frequency wave and miss the correct higher frequency one!

Original signal in red:

Sampling

In practice, then, we use the following sample rates.

• 16,000 Hz (samples/sec) Microphone (“Wideband”):• 8,000 Hz (samples/sec) Telephone

Why?• Need at least 2 samples per cycle• max measurable frequency is half sampling rate• Human speech < 10,000 Hz, so need max 20K• Telephone filtered at 4K, so 8K is enough

Quantization

• Quantization • Representing real value of

each amplitude as integer• 8-bit (-128 to 127) or 16-bit

(-32768 to 32767)• Formats:

• 16 bit PCM• 8 bit mu-law; log

compression• Byte order

• LSB (Intel) vs. MSB (Sun, Apple)

• Headers:• Raw (no header)• Microsoft wav• Sun .au

40 byteheader

WAV format

Waves have different frequencies

Time (s)0 0.02

–0.99

0.99

0

Time (s)0 0.02

–0.99

0.99

0

100 Hz

1000 Hz

Complex waves: Adding a 100 Hz and 1000 Hz wave

together

Time (s)0 0.05

–0.9654

0.99

0

Spectrum

100 1000Frequency in Hz

Am

plit

ude

Frequency components (100 and 1000 Hz) on x-axis

Spectra continued

• Fourier analysis: any wave can be represented as the (infinite) sum of sine waves of different frequencies (amplitude, phase)

Spectrum of one instant in an actual soundwave: many components across frequency range

Frequency (Hz)0 5000

0

20

40

Part of [ae] waveform from “had”

• Note complex wave repeating nine times in figure• Plus smaller waves which repeats 4 times for every large

pattern• Large wave has frequency of 250 Hz (9 times in .036 seconds)• Small wave roughly 4 times this, or roughly 1000 Hz• Two little tiny waves on top of peak of 1000 Hz waves

Back to spectrum

• Spectrum represents these freq components• Computed by Fourier transform, algorithm

which separates out each frequency component of wave.

• x-axis shows frequency, y-axis shows magnitude (in decibels, a log measure of amplitude)

• Peaks at 930 Hz, 1860 Hz, and 3020 Hz.

Spectrogram: spectrum + time dimension

FromMarkLiberman’sWeb site

Detecting Phones

• Two stages Feature extraction

Basically a slice of a spectrogram

Building a phone classifier (using GMM classifier)

MFCC: Mel-Frequency Cepstral Coefficients

Final Feature Vector

• 39 Features per 10 ms frame: 12 MFCC features 12 Delta MFCC features 12 Delta-Delta MFCC features 1 (log) frame energy 1 Delta (log) frame energy 1 Delta-Delta (log frame energy)

• So each frame represented by a 39D vector

Acoustic Modeling (= Phone detection)

• Given a 39-dimensional vector corresponding to the observation of one frame oi

• And given a phone q we want to detect

• Compute p(oi|q)

• Most popular method: GMM (Gaussian mixture models)

• Other methods Neural nets, CRFs, SVM, etc

Gaussian Mixture Models

• Also called “fully-continuous HMMs”• P(o|q) computed by a Gaussian:

p(o | q) 1

2exp(

(o )2

2 2)

Gaussians for Acoustic Modeling

• P(o|q):

P(o|q)

o

P(o|q) is highest here at mean

P(o|q is low here, very far from mean)

A Gaussian is parameterized by a mean and a variance:

Different means

Training Gaussians

• A (single) Gaussian is characterized by a mean and a variance

• Imagine that we had some training data in which each phone was labeled

• And imagine that we were just computing 1 single spectral value (real valued number) as our acoustic observation

• We could just compute the mean and variance from the data:

i 1

Tot

t1

T

s.t. ot is phone i

i2

1

T(ot

t1

T

i)2 s.t. ot is phone i

But we need 39 gaussians, not 1!

• The observation o is really a vector of length 39

• So need a vector of Gaussians:

p(o | q)

1

2D

2 2[d]d1

D

exp(

1

2

(o[d] [d])2

2[d]d1

D

)

Actually, mixture of gaussians

• Each phone is modeled by a sum of different gaussians

• Hence able to model complex facts about the data

Phone A

Phone B

Gaussians acoustic modeling

• Summary: each phone is represented by a GMM parameterized by M mixture weights M mean vectors M covariance matrices

Usually assume covariance matrix is diagonal

I.e. just keep separate variance for each cepstral feature

Where we are

• Given: A wave file• Goal: output a string of words• What we know: the acoustic model

How to turn the wavefile into a sequence of acoustic feature vectors, one every 10 ms

If we had a complete phonetic labeling of the training set, we know how to train a gaussian “phone detector” for each phone.

We also know how to represent each word as a sequence of phones

• What we knew from a few weeks ago: the language model• Next time:

Seeing all this back in the context of HMMs Search: how to combine the language model and the acoustic model to

produce a sequence of words

HMM for digit recognition task

Viterbi trellis for “five”

Viterbi trellis for “five”

Search space with bigrams

Viterbi trellis

75

Viterbi backtrace

76

Summary

• ASR Architecture• Phonetics Background• Five easy pieces of an ASR system

1) Lexicon2) Feature Extraction3) Acoustic Model (Phone detector)4) Language Model5) Viterbi decoding

A few advanced topics

Why foreign accents are hard

• A word by itself

• The word in context

Sentence Segmentation

• Binary classification task; judge the juncture between each two words:

• Features: Pause Duration of previous phone and rime Pitch change across boundary; pitch range of previous word

Disfluencies

• Reparandum: thing repaired• Interruption point (IP): where speaker

breaks off• Editing phase (edit terms): uh, I

mean, you know• Repair: fluent continuation

Example: Fragments: Incomplete or cut-off words:

Uh yeah, yeah, well, it- it- that’s right. And it-