IBM Research

Superhuman Speech Recognition:
Technology Challenges & Market Adoption David Nahamoo IBM Fellow Speech CTO, IBM Research July 2, 2008

© 2006 IBM Corporation

IBM Research

Overall Speech Market Opportunity
WW Voice-Driven Conversation Access Technology Forecast

•This chart represents all revenue for speech related ecosystem activity. •Revenue exceeded $1B for the 1st time in 2006 •Note also that hosted services will represent ½ of speech related revenue in 2011
2 © 2006 IBM Corporation
*Opus Research 02_2007

IBM Research

Speech Market Segments
Speech Segments Speech Self Service Speech Analytics Need-based Segmentation Transaction/Problem Solving Intelligence Market Usage Contact Centers Contact Centers, Media, Government Contact Centers, Government Media, Medical, Legal, Education, Government, Unified Messaging Contact Centers, Tourism, Global Digital Communities, Media (XCast) Embedded - Automotive, Mobile Devices, Appliances, Entertainment Mobile Internet, Yellow Pages SMS, IM, email

Speech Biometrics Speech Transcription Speech Translation

Security Information Access and provision Multilingual Communication

Speech Control Speech Search & Messaging

Command & Control Information Search & Retrieval

• Improved accuracy • Much larger vocabulary speech recognition system
3 © 2006 IBM Corporation

IBM Research

New Opportunity Areas
 Contact Centers Analytics
– Quality Assurance, Real Time Alerts, Compliance

 Media Transcription
– Closed captioning

 Accessibility
– Government, Lectures

 Content Analytics
– Audio-indexing, cross-lingual information retrieval, multi-media mining

 Dictation
– Medical, Legal, Insurance, Education

 Unified Communication
– Voicemail, Conference calls, email and SMS on hand held
4 © 2006 IBM Corporation

IBM Research

Target zone

Human Baseline for conversations

5

© 2006 IBM Corporation

IBM Research

Performance Results (2004 DARPA EARS Evaluation)
(Last public evaluation of English Telephony Transcription)

20 19 18 WER 17 16 15 14 0.1 1 xRT 10 100 IBM V3 V2 V1-V4 V4

IBM: Best Speed-Accuracy Tradeoff
6 © 2006 IBM Corporation

IBM Research

MALACH
Multilingual Access to Large Spoken ArCHives
• Funded by NSF, 5-year project (Started in Oct. 2001)

 Project Participants
– IBM, Visual History Foundation, Johns Hopkins University, University of Maryland, Charles University and University of West Bohemia

 Objective
– Improve access to large multilingual collections of spontaneous speech by advancing the state-of-the-art in technologies that work together to achieve this objective: Automatic Speech Recognition, Computer-Assisted Translation , Natural Language Processing and Information Retrieval

7

© 2006 IBM Corporation

IBM Research

MALACH: A challenging speech corpus
Multimedia digital archive: 116,000 hours of interviews with over 52,000 survivors, liberators, rescuers and witnesses of the Nazi Holocaust, recorded in 32 languages. Goal: improved access to large multilingual spoken archives

Challenges:
Disfluencies • A- a- a- a- band with on- our- on- our- arm Emotional speech • young man they ripped his teeth and beard out they beat him

Frequent interruptions:


8

CHURCH TWO DAYS these were the people who were to go to march TO MARCH and your brother smuggled himself SMUGGLED IN IN IN IN

© 2006 IBM Corporation

IBM Research

Effects of Customization (MALACH Data)
State-of-the-art ASR system trained on SWB data (8KHz)

Word Error Rates 90 80 70 60 50 40 30 20 Jan. '02 Oct. '02 Oct. '03 June '04 Nov. '04

MALACH Training data seen by AM and LM

fMPE, MPE, Consensus decoding

9

© 2006 IBM Corporation

IBM Research

Improvement in Word Error Rate for IBM embedded ViaVoice
6 5 4 3 2 1 0 WER across 3 car speeds and 4 grammars

10

© 2006 IBM Corporation

IBM Research

Progress in Word Error Rate – IBM WebSphere Voice Server
Grammar Tasks over Telephone 2001 - 2006
6 5 4
Word Error 3 Rate % 2

1 0
WVS

45% relative improvement in WER the last 2.5 years 20% relative improvement in speed in the last 1.5 years

11

© 2006 IBM Corporation

IBM Research

Multi-Talker Speech Separation Task
Lay white at X 8 soon male and female speaker at 0dB

Bin Green with F 7 now

12

© 2006 IBM Corporation

IBM Research

Two Talker Speech Separation Challenge Results
Examples:
Mixture

13

Recognition Error

© 2006 IBM Corporation

IBM Research

Comparison of Human & Machine Speech Recognition
100

Voicemail SWITCHBOARD 10 BROADCAST NEWS BROADCAST-HUMAN SWITCHBOARD-HUMAN

1 1992 1993
70

1994

1995

1996

1997

1998

1999

2000

Word Error Rate

60 50 40 30 20 10 0 WSJ Broadcast Conv Tel Vmail SWB Call center Meeting

Clean Speech Human-Machine
14

Spontaneous Speech Human-Human
© 2006 IBM Corporation

IBM Research

IBM’s Superhuman Speech Recognition
Universal Recognizer
• Any accent • Any topic • Any noise conditions • Broadcast, phone, in car, or live • Multiple languages • Conversational

15

© 2006 IBM Corporation

IBM Research

Human Experiments

Question: – Can post-processing of recognizer hypotheses by humans improve accuracy? – What is the relative contribution of linguistic vs. acoustic information (in this postprocessing operation?) Experiment – Produce recognizer hypotheses in form of “sausages” – Allow human to correct output either with linguistic information alone or with short segments of acoustic information Results – Human performance still far from maximum possible, given information in “sausages” – Recognizer hypothesized linguistic context information not useful by itself – Acoustic information in limited span (1 sec. average) marginally useful What we learned – Hard-to-design – Expensive to conduct – Hard to decide if not valuable

that it I

could cuts

stem

and

down

on

comes they
16

stay them

I’m

cut

© 2006 IBM Corporation

IBM Research

Acoustic Modeling Today
 Approach: Hidden Markov Models
– Observation densities (GMM) for P( feature | class ) • Mature mathematical framework, easy to combine with linguistic information • However, does not directly model what we want i.e., P( words | acoustics )

 Training: Use transcribed speech data
– Maximum Likelihood – Various discriminative criteria

 Handling Training/Test Mismatches:
– Avoid mismatches by collecting “custom” data – Adaptation & adaptive training algorithms

 Significantly worse than humans for tasks with little or no linguistic
information - e.g., digits/letters recognition

 Human performance extremely robust to acoustic variations
– due to speaker, speaking style, microphone, channel, noise, accent, & dialect variations

 Steady progress over the years
17

Continued progress using current methodology very likely in the © 2006 IBM Corporation future

IBM Research

Towards a Non-Parametric Approach to Acoustics
 General Idea: Back to pattern recognition basics! – Break test utterance into sequence of larger segments (phone, syllable, word, phrase) – Match segments to closest ones in training corpus using some metric (possibly using long distance models) – Helps to get it right if you’ve heard it before  Why prefer this approach over HMMs? – HMMs compress training by x1000; too many modeling assumptions
• • 1000hrs ~ 30Gb; State-of-the-art acoustic models ~ 30Mb Relaxing assumptions have been key to all recent improvements in acoustic modeling

 How can we accomplish this?
– Store & index training data for rapid access of training segments close to test segments – Develop a metric D( train_seq, test_seq): obvious candidate is DTW with appropriate metric and warping rules

 Back to the Future? – Reminiscent of DTW & Segmental models from late 80’s – ME was missing – Limited by computational resources (storage/cpu/data) then & so HMMs won  Implications: – Need 100x more data for handling larger units (hence 100x more computing resources) – Better performance with more data – likely to have “heard it before”
18 © 2006 IBM Corporation

IBM Research

Utilizing Linguistic Information in ASR
 Today’s standard ASR does not explicitly use linguistic information
– But recent work at JHU, SRI and IBM all show promise – Semantic structured LM improves ASR significantly for limited domains  Reduces WER by 25% across many tasks (Air Travel, Medical)

 A large amount of linguistic knowledge sources now available, but not used
for ASR
 Inside IBM  WWW text: Raw text: 50 million pages ~25 billion words, ~10% useful after cleanup  News text: 3-4 billion words, broadcast or newswires  Name entity annotated text: 2 million words tagged  Ontologies  Linguistic knowledge used in rule-based MT system  External  WordNet, FrameNet, Cyc ontologies  PennTreeBank, Brown corpus (syntactic & semantic annotated)  Online dictionaries and thesaurus  Google

19

© 2006 IBM Corporation

IBM Research

Super Structured LM for LVCSR
Cohe nce re : se antic,syntactic, m pragm atic
Wash your clothe with s soap/ soup. David and his/ he fathe r r walke into the room d . I ate a one nine pound / ste ak.

Dialogue State

Semantic Parser

World Knowledge Named Entity Embedde d Grammar Word Class

Documen t Type Speaker (turn, gender, ID)

W , ..., W
1

N

Syntactic Parser

•Acoustic Confusability: LM should be optimized to distinguish between acoustic confusable sets, rather than based on N-gram counts •Automatic LM adaptation at different levels: discourse, semantic structure, and phrase levels
20 © 2006 IBM Corporation

IBM Research

Combination Decoders
 “ROVER” is used in all current systems
– NIST tool that combines multiple system outputs through voting

 Individual systems currently designed in an ad-hoc manner  Only 5 or so systems possible
“I feel shine today” “I veal fine today”

“I feel fine toady”

“I feel fine today”

An army (“Million”) of simple decoders • Each makes uncorrelated errors
21 © 2006 IBM Corporation

IBM Research

Million Feature Paradigm: Acoustic information for ASR

Segmental analysis Broadband features Narrowband features

Trajectory features Onset features

Discard transient noise; Global adaptation for stationary noise Noise Sources

Information Sources

• Feature definition is key challenge • Maximum entropy model used to compute word probabilities. • Information sources combined in unified theoretical framework. • Long-span segmental analysis inherently robust to both stationary and transient noise
22 © 2006 IBM Corporation

IBM Research

Implications of the data-driven learning paradigm
 ASR systems give the best results when test data is similar to the
training data

 Performance degrades as the test data diverges from the training data
– 1. 2. Differences can occur both at the acoustic and linguistic levels, e.g. A system designed to transcribe standard telephone audio (8kHz) cannot transcribe compressed telephony archives (6kHz) A system designed for a given domain (e.g. broadcast news) will perform worse on a different domain (e.g. dictation)

 Hence the training and test sets have to be carefully chosen if the task
at hand expects a variety of acoustic sources

23

© 2006 IBM Corporation

IBM Research

Generalization Dilemma
Want to get here:

Correct complex model
(simple model on the right manifold)

Performance

Model combination: Can we at least get best of both worlds? Simple model

Complex model: brute force learning

The Gutter of Data Addiction

In-Domain Test Conditions
24

Out-of-Domain

© 2006 IBM Corporation

IBM Research

Summary
 Continue the current tried-and-true technical approach  Continue the yearly milestones and evaluations  Continue the focus on accuracy, robustness, & efficiency  Increase the focus on quantum leap innovation  Increase the focus on language modeling  Plan for 2 orders of magnitude increase in
– Access to annotated speech and text data – Computing resources

 Improve cross-fertilization among different projects

25

© 2006 IBM Corporation