Characterizing and RecognizingSpoken Corrections in

Human-Computer Dialog

Gina-Anne LevowAugust 25, 1998

Error Correction Spiral U: Give me the price for AT&T. S: Hewlett-Packard was 83 ¾, up 2 ½ since yesterday. U: Give me the price for AT&T. S: Data General was at 10 ½ up a quarter. U: Give me the price for AT&T. S: Hewlett-Packard was 83 ¾, up 2 ½. U: Give me the price for American Telephone and

Telegraph. S: Excuse me?

Identifying Corrections

Most "Reasoning" Approaches Correction => Violates Task, Belief Constraint Requires Tight Task, Belief Model Often Requires Accurate Recognition

This Approach Uses Acoustic or Lexical Information Content, Context Independent

Accomplishments

Corrections vs Original Inputs Significant Differences: Duration, Pause, Pitch

Corrections vs Recognizer Models Contrasts: Phonology and Duration

Correction Recognition Decision Tree Classifier: 65-77% accuracy

Human Baseline ~80%

Why Corrections?

Recognizer Error Rates ~25% REAL meaning of utterance

user intent Corrections misrecognized 2.5X as often Hard to Correct => Poor Quality System

Why it's Necessary

Error Repair Requires Detection Errors can be very difficult to detect

E.g. Misrecognitions Focus Repair Efforts Corrections Decrease Recognition Accuracy

Adaptation Requires Identification

Why is it Hard?

Recognition Failures and Errors Repetition <> Correction

500 Strings => 6700 Instances (80%) Speech Recognition Technology

Variation - Undesirable, Suppressed

Corrections Within

Roadmap

Data Collection and Description SpeechActs System & Field Trial

Characterizing Corrections Original-Repeat Pair Data Analysis Acoustic and Phonological Measures & Results

Recognizing Corrections Conclusions and Future Work

SpeechActs System

Speech-Only System over the Telephone Access to Common Desktop Applications

Email, Calendar, Weather, Stock Quotes BBN's Hark Speech Recognition, Centigram

TruVoice Speech Synthesis In-house: Natural Language Analysis

Back-end Applications, Dialog Manager

Example Interaction SpeechActs: You have 14 new messages and 47 old

messages. The first new message is from Eric Baatz regarding "report completed".

User: Let me hear it. SpeechActs: "The first draft is ready for your comments." User: Reply and Include the current message. SpeechActs: Preparing message to Eric Baatz. Begin

recording after the tone. When finished, pause for several seconds.

User: Eric, I'll get those to you this afternoon. SpeechActs: Say cancel, send or review.

Example Interaction: Con't

User: Send. SpeechActs: Message Sent. What now? User: Next message. SpeechActs: New Message two from Stuart Adams, subject

"Can we meet at 3:00 today?" User: Switch to Calendar.

System Data Overview

Approximately 60 hours of interactions Digitized at 8kHz, 8-bit mu-law encoding

18 subjects: 14 novices, 4 experts, single shots 7529 user utterances, 1961 errors ~25% error P(error | correct) = 18%; P(error | error) = 44%

System: Recognition Error Types

Rejection Errors - Below Recognition Level U: Switch to Weather S (Heard): <nothing> S (said): Huh?

Misrecognition Errors - Substitution in Text U: Switch to Weather S (Heard): Switch to Calendar S (Said): On Tuesday August 25, you have defense

1250 Rejections ~2/3 706 Misrecognitions ~1/3

Roadmap

Data Collection and Description SpeechActs System & Field Trial

Characterizing Corrections Original-Repeat Pair Data Analysis Acoustic and Phonological Measures & Results Divergence from Recognizer Models

Recognizing Corrections Conclusions and Future Work

Analysis: Data

300 Original Input-Repeat Correction Pairs Lexically Matched, Same Speaker Example:

S: (Said): Please say mail, calendar, weather. U: Switch to Weather. Original S (Said): Huh? U: Switch to Weather. Repeat.

Analysis: Duration

Automatic Forced Alignment, Hand-Edited Total: Speech Onset to End of Utterance Speech: Total - Internal Silence Contrasts: Original Input/Repeat Correction

Total: Increases 12.5% on average Speech: Increases 9% on average

Analysis: Pause

Utterance Internal Silence > 10ms Not Preceding Unvoiced Stops(t), Affricates(ch)

Contrasts: Original Input/Repeat Correction Absolute: 46% Increase Ratio of Silence to Total Duration: 58% Increase

Pitch Tracks

Analysis: Pitch I

ESPS/Waves+ Pitch Tracker, Hand-Edited Normalized Per-Subject:

(Value-Subject Mean) / (Subject Std Dev) Pitch Maximum, Minimum, Range

Whole Utterance & Last Word Contrasts: Original Input/Repeat Correction

Significant Decrease in Pitch Minimum Whole Utterance & Last Word

Analysis: Pitch II

Analysis: Pitch III

Internal Pitch Contours: Pitch Accent Steepest Rise, Steepest Fall, Slope Sum Overall => Not Significant Misrecognitions Only: Original vs Repeat

Significant Increases: Steepest Rise, Slope Sum

Pitch Contour Detail

Exclude Boundary Tone Region 5-Point median smoothing (Taylor 1996) Piecewise linear contour between max and min

Analysis: Overview

Significant Differences: Original/Correction Duration & Pause

Significant Increases: Original vs Correction Pitch

Significant Decrease in Pitch Minimum Increase in Final Falling Contours Misrecognitions: Increase in Pitch Variability

Conversational-to-Clear Speech Shift Contrastive Use of Pitch Accent

Roadmap

Data Collection and Description SpeechActs System & Field Trial

Characterizing Corrections Original-Repeat Pair Data Analysis Acoustic and Phonological Measures & Results Divergence from Recognizer Models

Recognizing Corrections Conclusions and Future Work

Analysis: Phonology

Reduced Form => Citation Form Schwa to unreduced vowel (~20)

E.g. Switch t' mail => Switch to mail. Unreleased or Flapped 't' => Released 't' (~50)

E.g. Read message tweny => Read message twenty Citation Form => Hyperclear Form

Vowel or Syllabic Insertion (~20) E.g. Goodbye => Goodba-aye

Analysis: Overview II

Original vs Correction & Recognizer Model Phonology

Reduced Form => Citation Form => Hyperclear Form Conversational to (Hyper) Clear Shift

Duration Contrast between Final and Non-final Words Departure from ASR Model

Increase for Corrections, especially Final Words

Roadmap

Data Collection and Description SpeechActs System & Field Trial

Characterizing Corrections Original-Repeat Pair Data Analysis Acoustic and Phonological Measures & Results Divergence from Recognizer Models

Recognizing Corrections Conclusions and Future Work

Learning Method Options (K)-Nearest Neighbor

Need Commensurable Attribute Values Sensitive to Irrelevant Attributes Labeling Speed - Training Set Size

Neural Nets Hard to Interpret Can Require More Computation & Training Data +Fast, Accurate when Trained

Decision Trees Intelligible, Robust to Irrelevant Attributes +Fast, Compact when Trained ?Rectangular Decision Boundaries, Don't Test Feature Combinations

Alternative: Mixture of Experts

Learning Method Options (K)-Nearest Neighbor

Need Commensurable Attribute Values Sensitive to Irrelevant Attributes Labeling Speed - Training Set Size

Neural Nets Hard to Interpret Can Require More Computation & Training Data +Fast, Accurate when Trained

Decision Trees <= Intelligible, Robust to Irrelevant Attributes +Fast, Compact when Trained ?Rectangular Decision Boundaries, Don't Test Feature Combinations

Decision Tree Features 38 Features Total, E.g.

15 for best trees Pause

Total Pause Duration Pause / Total Duration

Duration Total Duration (uttdur) Speaking Rate (sps) Normalized Duration

Amplitude Max, Mean, Last Max-Last (ampdiff) Mean-Last (ampdelta)

Pitch Max, Min, Range

Global, Last Word Range/Total Contour Max, min, sum slope

Decision Tree Training & Testing

Data: 50% Original Inputs, 50% Repeat Corrections

Classifier Labels: Original, Correction 7-Way Cross-Validation

Train on 6/7 of data, Test on remaining 1/7 Subsets drawn at random according to distribution Cycle through all subsets, training & testing Report average results on unseen test data

Recognizer: Results (Overall)

Tree Size: 57 (unpruned), 37 (pruned) Minimum of 10 nodes per branch required

First Split: Normalized Duration (All Trees) Most Important Features:

Normalized & Absolute Duration, Speaking Rate 65% Accuracy - Null Baseline-50%

Example Tree

Classifier Results: Misrecognitions

Most important features: Absolute and Normalized Duration Pitch Minimum and Pitch Slope

77% accuracy (with text) 65% (acoustic features only)

Null baseline - 50% Human baseline - 79.4% (Hauptman & Rudnicky 1990)

Classifier Results: Misrecognitions

Most important features: Absolute and Normalized Duration Pitch Minimum and Pitch Slope

77% accuracy (with text) 65% (acoustic features only)

Errors, most trees: ½ false positive, ½ false negative Null baseline - 50% Human baseline - 79.4% (Hauptman & Rudnicky 1990)

Misrecognition Classifier

Roadmap

Data Collection and Description Characterizing Corrections Recognizing Corrections

Conclusions and Future Work

Accomplishments

Contrasts between Originals vs Corrections Significant Differences in Duration, Pause, Pitch

Conversational-to-Clear Speech Shifts Shifts away from Recognizer Models Corrections Recognized at 65-77%

Near-human Levels

The Recipe

Original/Correction Training Set (300+ sets) Labeled, Transcribed, Digitized, Corpus or Wizard

Acoustic Analyses Pitch Tracking, Silence Detection, Speaking Rate,...

Classifier Training & Tuning Confidence Measure (Weighted Pessimistic Error)

Phonological Rule Extraction Durational Contrast Modeling Repair Dialog Management

Future Work

Modify ASR Duration Model for Correction Reflect Phonological and Duration Change

Identify Locus of Correction for Misrecognitions Preliminary tests:

26/28 Corrected Words Detected, 2 False Alarms