Intonation and Multi-Language Scenarios Andrew Rosenberg Candidacy Exam Presentation October 4, 2006 Talk Overview Use and Meaning of Intonation Automatic Analysis of Intonation “Multi-Language Scenarios” Second Language Learning Systems Speech-to-Speech Translation 10/04/2006 2 Use and Meaning of Intonation Why do multi-language scenarios need intonation? Intonation indicates focus and contrast Intonation disambiguates meaning Intonation indicates how language is being used Discourse structure, Speech acts, Paralinguistics 10/04/2006 3 Examples of Intonational Features ToBI Examples Categorical Features Pitch Accent H* - Mariana(H*) won it L+H* - Mariana(L+H*) won it L* - Will you have marmalade(L*) or jam(L*) Phrase Boundaries Intermediate Phrase Boundary (3) Intonational Phrase Boundary (4) Oh I don’t know (4) it’s got oregano (3) and marjoram (3) and some fresh basil (4) Continuous Features 10/04/2006 Pitch Intensity Duration 4 Use and Meaning of Intonation Paper List Emphasis Accent is Predictable (If You’re a Mind Reader) Bolinger, 1972 Prosodic Analysis and the Given/New Distinction Brown, 1983 The Prosody of Questions in Natural Discourse Hedberg and Sosa, 2002 Syntax The Use of Prosody in Syntactic Disambiguation Price, et al., 1991 Discourse Structure Prosodic Analysis of Discourse Segments in Direction Giving Monologues Nakatani and Hirschberg, 1996 Paralinguistics Acoustic Correlates of Emotion Dimension in View of Speech Synthesis Schröder, et al., 2001 10/04/2006 5 Accent is Predictable (If You're a Mind Reader) Dwight Bolinger, 1972 Harvard University Nuclear Stress Rule Stress is assigned to the rightmost stress-able vowel in a major constituent (Chomsky and Halle 1968) “Once the speaker has selected a sentence with a particular syntactic structure and certain lexical items...the choice of stress contour is not a matter subject to further independent decision” Selected Counterexamples to NSR Coordinated Infinitives can be accented or not Terminal prepositions are rarely accented Why are you coming indoors? -- I’m coming indoors because the sun is shining Predictable or less semantically rich items are less likely to be accented 10/04/2006 I need a light to read by Focus v. Topic v. Comment I have a clock to clean and oil v. I still have most of the garden to weed and fertilize I have a point to make v. I have a point to emphasize I’ve got to go see a guy v. I’ve got to go see a friend [semi-pronouns?] 6 Prosodic Analysis and the Given/New Distinction Gillian Brown, 1983 Information Structure of Discourse Entities (Prince 1981) Experiment Given (or Evoked) Information: “recoverable either anaphorically or situational” (Halliday,1967) New Information: “non recoverable...” Inferable Information: e.g. driver is inferable given bus One subject was asked to describe a diagram to another who would reproduce it. Entities are marked as new/brand-new, new/inferred, evoked/context (pen, paper, etc.), evoked/current (most recent mention) or evoked/displaced (previously mentioned) Prominence realizations 10/04/2006 87% of new/brand-new and 79% of new/inferred entities 2% of evoked/context, 0% of evoked current, 4% of evoked/displaced 7 The Prosody of Questions in Natural Discourse Nancy Hedberg and Juan Sosa, 2002 Simon Fraser University Accenting behavior in question types Wh-Questions (whq) vs. Yes/No Questions (ynq) in spontaneous speech from “McLaughlin Group” and “Washington Week” The “locus of interrogation” Either Wh-word or Fronted Auxiliary Verb Wh-words are often accented with L+H* and rarely deaccented Ynqs show no consistent accenting behavior Whqs are produced with falling intonation 80% of the time Only 34% of Ynqs are produced with rising intonation Topic Pitch Accent 10/04/2006 70.5% of positive ynqs deaccent 88% of negative ynqs use L+H* Nuclear Tune Where are you? Do you like pie? The topic of both whqs and ynqs are less often accented with L+H* than the locus of interrogation 8 The Use of Prosody in Syntactic Disambiguation Patti Price1, Mari Ostendorf2, Stefanie Shattuck-Hufnagel3, Cynthia Fong2, 1991 1SRI, 2Boston University, 3MIT Relationship between syntax and intonation. Methodology 7 Types of syntactically ambiguous sentences spoken by 4 professional radio announcers Ambiguous sentences were produced within disambiguating paragraphs. The speakers were not informed of the sentence of interest and only produced one version per session. Subjects selected the more appropriate surrounding context. Subjects only rated one version per session. Analysis Manual labelling of phrase breaks and accents [not ToBI] Phrase breaks and their relative size differentiate the two versions. 10/04/2006 Characterized by lengthening, pauses and boundary tone 9 Example Syntactic ambiguities Parentheticals v. non-parenthetical clause Apposition v. attached NP [Only one remembered,][the lady in red] [Only one remembered the lady in red] Main clauses w/ coordinating conjunction v. main and subordinate clause [Mary knows many languages,][you know] [Mary knows many languages (that) you (also) know] [Jane rides in the van][and Ella runs] [Jane rides in the van Ann Della runs] Tag question v. attached NP 10/04/2006 [Mary and I don’t believe][do we?] [Mary and I don’t believe Dewey.] 10 Example Syntactic ambiguities Far v. Near attachment of final phrase Left v. Right attachment of middle phrase [Raoul murdered the man][with the gun] (Raoul has a gun) [Raoul murdered [the man with the gun]] (the man has a gun) [When you learn gradually][you worry more] [When you learn][gradually you worry more] Particles v. Prepositions 10/04/2006 [They may wear down the road] (the treads hurt the road) [They may wear][down the road] (the treads erode) 11 Prosodic Analysis of Discourse Segments in Direction Giving Monologues Julia Hirschberg1 and Christine Nakatani2,1996 1AT&T Labs, 2Harvard University Intonation is used to indicate discourse structure Boston Directions Corpus Is a speaker beginning a topic? ending one? Entails a broader information (linguistic, attentional, intentional) structure than given/new entity status. Manual ToBI annotation, and Discourse segmentation Acoustic-prosodic correlates of discourse segment initial, medial and final phrases. Segment Initial v. non-initial Segment Medial v. Final 10/04/2006 Higher max, mean F0 and Energy. Longer preceding and shorter following pauses Increases in F0 and Energy from previous phrase Medial has a slower speaking rate and shorter subsequent pause Relative increase in F0 and Energy from previous phrase 12 BDC Discourse structure example [describe green line portion of journey] and get on the Green Line [describe direction to take on green line] we will take the Green Line south toward Park Street [describe which green line to take (any)] we can get on any of the Green Lines at Government Center and take them south to Park Street [describe getting off the green line] once we are at Park Street we will get off [describe red line portion of journey] and get on the red line of the T 10/04/2006 13 Acoustic Correlates of Emotion Dimension in View of Speech Synthesis Marc Schröder1, Roddy Cowie2, Ellen Douglas-Cowie2, Machiel Westerdijk3, Stan Gielen3, 2001 1University of Saarland, 2Queen’s University, 3University of Nijmegen Paralinguistic Information That information that is transmitted via language that is not strictly “linguistic”. E.g. emotion, humor, charisma, deception Emotional Dimensions Activation - Degree of readiness to act Evaluation - Positive v. Negative Power - Dominance v. Submission For example, Happiness - High Activation, High Evaluation, High Power Anger - High Activation, Low Evaluation, Low Power Sadness - Low Activation, Low Evaluation, Very Low Power 10/04/2006 14 Acoustic Correlates to Emotion Dimension ctd. Manual annotation of emotional content of spontaneous speech from 100 speakers in activation-evaluation-power space. High Activation strongly correlates High F0 mean and range, longer phrases, shorter pauses, large and fast F0 rise and fall, increased intensity, flat spectral slope Negative Evaluation correlates Fast F0 falls, long pauses, increased intensity, more pronounced intensity maxima High Power correlates 10/04/2006 Low F0 mean, (female) shallow F0 rise and falls, reduced intensity (male) increased intensity 15 Use and Meaning Of Intonation Summary Intonation can provide information about: Focus Contrast I want the red pen (...not the blue one) 10/04/2006 Information Status (given/new) Speech Acts Discourse Status Syntax Paralinguistics 16 Automatic Analysis of Intonation How can the information transmitted via intonation be understood computationally? What computational techniques are available? How much human annotation is needed? 10/04/2006 17 Automatic Analysis of Intonation Paper List (1/2) Supervised Methods Automatic Recognition of Intonational Features Wightman and Ostendorf, 1992 An Automatic Prosody Recognizer Using a Coupled Multi-Stream Acoustic Model and a Syntactic-Prosodic Language Model Ananthakrishnan and Narayanan, 2005 Perceptually-related Acoustic-Prosodic Features of Phrase Finals in Spontaneous Speech Ishi, et al., 2003 Alternate ways of representing Intonation Direct Modeling of Prosody: An Overview of Applications in Automatic Speech Processing Shriberg and Stolcke 2004 The Tilt Intonation Model Taylor, 1998 10/04/2006 18 Automatic Analysis of Intonation Paper List (2/2) Unsupervised Methods Unsupervised and Semi-supervised Learning of Tone and Pitch Accent Levow, 2006 Reliable Prominence Identification in English Spontaneous Speech Tamburini, 2006 Feature Analysis Spectral Emphasis as an Additional Source of Information in Accent Detection Heldner, 2001 Duration Features in Prosodic Classification: Why Normalization Comes Second, and what they Really Encode. Batliner, et al., 2001 10/04/2006 19 Supervised Methods Require annotated data Pitch Accent and Phrase Boundaries are the two main prosodic events that are detected and classified 10/04/2006 20 Automatic Recognition of Intonational Features Colin Wightman and Mari Ostendorf, 1992 Boston University Detection of boundary tones and pitch accents on syllables Decision tree-based acoustic quantization for use with an HMM Four-way classification {Pitch Accent, Boundary Tone, Both, Neither} Features Is the syllable lexically stressed? F0 contour representation Max, min F0 context normalization Duration Pause information Mean energy Results 10/04/2006 Prominence: Correct 86% False alarm 14% Boundary Tone: Correct 77% False alarm 3% 21 An Automatic Prosody Recognizer Using a Coupled MultiStream Acoustic Model and a Syntactic-Prosodic Language Model Shankar Aranthakrishnan and Shrikanth Narayanan, 2005 University of Sothern California There are three asynchronous information streams that contribute to intonation Coupled HMM trained on 1 hour of radio news speaker data with ASR hypotheses and POS tags Tag syllable as long/short, stressed/unstressed, boundary/non-boundary Includes language model relating POS and prosodic events Syntax alone provides the best results for boundary tone detection: Pitch - duration and distance from mean of piecewise linear fit of f0 Energy - frame level intensity normalized w.r.t utterance Duration - normalized vowel duration of the current syllable and following pause duration Correct 82.1% False Alarm 12.93% Stress detection false alarm rate is nearly halved by inclusion of acoustic information 10/04/2006 Syntax alone: 79.7% / 22.25% Syntax + acoustics: 79.5% / 13.21% 22 Perceptually-related Acoustic-Prosodic Features of Phrase Finals in Spontaneous Speech Carlos Toshinori Ishi, Parham Mokhtari, Nick Campbell, 2003 ATR/Human Information Science Labs Phrase-final behavior can indicate speech act, certainty, discourse/topic structure, etc. Classification of phrase-final behavior in Japanese Pitch features Mean F0 of first and second half of phrase final Pitch target of first and second half Min, max, (pseudo-) slope, reset of phrase final Using a classification tree, 11 tone classes could be classified with 75.9% accuracy Majority class baseline: 19.6% 10/04/2006 23 Perceptually-related Acoustic-Prosodic Features of Phrase Finals in Spontaneous Speech Carlos Toshinori Ishi, Parham Mokhtari, Nick Campbell, 2003 ATR/Human Information Science Labs Tone Type Perceptual Properties (Hattori 2002) X-JToBI BPM 1a Low L% 1b Low+Falling tone L% 1bE Low+Falling+Extended L% 2a High L%+H% 2aA High+Aspirated L%+H% 2b High+Lengthened L%+H%> 2c Low+Rising tine L%+LH% 2cE Low+Rising+Extended L%+LH% 2cS Low+Rising+Short L%+LH% 3 High+Falling tone L%+HL% 5 High+Fall-Rise tone L%+HLH% 10/04/2006 24 Direct Modeling of Prosody: An Overview of Applications in Automatic Speech Processing Elizabeth Shriberg and Andreas Stolcke, 2004 SRI, ICSI Do we need to explicitly model prosodic features? Why not provide acoustic/prosodic information directly to other statistical models? Task-based integration of features and models Event Language Model Augment a typical n-gram language model with prosodic event classes Event Prosodic Model Grow decision trees or use GMMs to generate P(Event|Signal) Continuous Prosodic Features 10/04/2006 Duration from ASR Pitch, energy, voicing normalizations and stylizations Task specific features: e.g. Number of repeat attempts 25 Direct Modeling of Prosody Tasks Structural Tagging Sentence/topic boundary and disfluency (interruption point) detection Uses Language Model + Event Prosodic Model Sentence boundary results Telephone: accuracy improved 7% BN: 19% error reduction Pragmatics/Paralinguistics Dialog act classification and frustration detection Uses Language Model and Dialog “grammar” + Event Prosodic Model Results: Statement v. Question 16% error reduction Agreement v. Backchannel 16% error reduction Frustration 27% error reduction (using “repeated attempt” feature) 10/04/2006 26 Direct Modeling of Prosody Tasks Speaker Recognition Typical approaches use spectral information Use Continuous Prosodic Features Including phone duration features can reduce error by 50% Word Recognition Words can be recognized simultaneously with prosodic events (Event Language Model) Spectral and prosodic information can be used to model word hypotheses Phone duration, pause information along with sentence and disfluency detection reduces error by 3.1% 10/04/2006 27 The Tilt Intonation Model Paul Taylor, 1998 University of Edinburgh The Tilt Model describes accent and boundary tones as “intonational events” characterized by pitch movement Events (accent, boundary, neither, silence) are automatically detected using an HMM with pitch, energy, and first and second order difference of both Accuracy ranged from 35%-47% with correct identification of events between 60.7% and 72.7% Tilt parameter was then extracted from the HMM hypotheses. F0 synthesis with machine and human derived Tilt parameters differed by < 1Hz rmse on DCEIM test set 10/04/2006 28 Tilt parameter tilt Arise A fall 2(Arise A fall ) Drise Dfall 2(Drise Dfall ) 10/04/2006 29 Unsupervised Models of Intonation Annotating Intonation is expensive 100x real time for full ToBI labeling Human Annotations are errorful Human agreement ranges from 80-90% Unsupervised Methods are 1. Inexpensive Data doesn’t require manual annotation 2. Consistent Performance is not reliant on human consistency 10/04/2006 30 Unsupervised and Semi-supervised Learning of Tone and Pitch Accent Gina-Anne Levow, 2006 University of Chicago What can we do without gold-standard data Also, does Lexical Tone in Mandarin Chinese vary in similar dimensions as Pitch Accent? Semi- and Unsupervised speaker-dependent clustering into 4 accent classes (unaccented, high, low, downstepped) Forced alignment-based syllable Features: Speaker normalized f0, f0 slope and intensity Context: prev. following syllables values and first order differences Semi-supervised: Laplacian Support Vector Machines Tone (clean speech): 94% accuracy (99% supervised) Pitch Accent (2-way): 81.5% accuracy (84% supervised) Unsupervised: k-means clustering, Asymmetric k-lines clustering Tone (clean speech): 77% accuracy Pitch Accent (4-way): 78.4% accuracy (80.1% supervised) 10/04/2006 31 Reliable Prominence Identification in English Spontaneous Speech Fabio Tamburini, 2005 University of Bologna Unsupervised metric to describe the prominence of a syllable Calculated over the nucleus Prom = en500-4000 + dur + enov (Aevent + Devent) High spectrum energy Duration Full spectrum energy Tilt parameters (f0 amplitude and duration) By tuning a threshold, 18.64% syllable error rate on TIMIT 10/04/2006 32 Feature Analysis Intonation is generally assumed to be realized as a modification of Pitch Energy Duration How do each of these contribute to realization of specific prosodic events? 10/04/2006 33 Spectral Emphasis as an Additional Source of Information in Accent Detection Mattias Heldner, 2001 Umeå University Close inspection of spectral emphasis as discriminating accented and non-accented syllables in read Swedish Spectral emphasis: difference (in dB) of energy in the first formant and full spectrum First formant energy was extracted using a dynamic low pass filter with a cut off that followed f0 Classifier: The word in a phrase with the highest spectral emphasis/intensity/pitch is “focally accented”. Results: 10/04/2006 Spectral Emphasis: 75% correct Overall Intensity: 69% correct Pitch peak: 67% correct 34 Duration Features in Prosodic Classification: Why Normalization Comes Second, and what they Really Encode Anton Batliner, Elmar Nöth, Jan Buckow, Richard Huber, Volker Warnke, Heinrich Niemann, 2001 University of Erlangen-Nuremberg When vowels are stressed, accented or phrase-final they tend to be lengthened. What’s the best way to measure the duration of a word? Duration is normalized in three ways DURNORM - normalized w.r.t. ‘expected’ duration Expected duration calculated by the mean and std.dev. of a vowel scaled by a ROS approximation. DURSYLL - normalized w.r.t. number of syllables DURABS - raw duration In both German and English on boundary and accent tasks, DURABS classified the best followed by DURSYLL followed by DURNORM Duration inadvertently encodes semantic information 10/04/2006 Complex words tend to have more syllables and tend to be accented more frequently; common words (particles, backchannels) tend to be shorter DURNORM and DURSYLL are able to classify well (if worse) despite obfuscating this information 35 Automatic Analysis of Intonation Summary Various of models can be used to analyze both pitch accents and phrase boundaries: Supervised Direct Discriminative modelling Semi- and unsupervised learning Research has also examined how accents and phrase breaks are realized in a constrained acoustic dimensions 10/04/2006 36 Second Language Learning Systems Automated systems can be used to improve pronunciation and intonation of second language learners. Native intonation is rarely emphasized in classrooms and is often the last thing nonnative speakers learn. Focus will be more on computational approaches (diagnosis, evaluation) over pedagogical concerns 10/04/2006 37 Second Language Learning Systems Paper List (1/2) Pronunciation Evaluation The SRI EduSpeakTM System: Recognition and Pronunciation Scoring for Language Learning Automatic Localization and Diagnosis of Pronunciation Errors for Second-Language Learners of English Herron, et al., 1999 Automatic Syllable Stress Detection Using Prosodic Features for Pronunciation Evaluation of Language Learners 10/04/2006 Franco, et al., 2000 Tepperman and Narayanan, 2005 38 Second Language Learning Systems Paper List (2/2) Fluency, Nativeness and Intonation Evaluation A Visual Display for the Teaching of Intonation Quantitative Assessment of Second Language Learner’s Fluency: An Automatic Approach Imoto, et al., 2002 A study of sentence stress production in Mandarin speakers of American English 10/04/2006 Teixeira, et al., 2000 Modeling and Automatic Detection of English Sentence Stress for Computer-Assisted English Prosody Learning System Cucchiarini, et al., 2002 Prosodic Features for Automatic Text-Independent Evaluation of Degree of Nativeness for Language Learners Spaai and Hermes, 1993 Chen, et al., 2001 39 Pronunciation Evaluation The segmental context and lexical stress of a production determines whether it is pronounced correctly or not. 10/04/2006 40 The SRI EduSpeakTM System: Recognition and Pronunciation Scoring for Language Learning Horacio Franco, Victor Abrash, Kristin Precoda, Harry Bratt, Ramana Rao, John Butzberger, Romain Rossier, Federico Cesari, 2000 SRI Recognition Non-native speech recognition is errorful. A native HMM recognizer was adapted to non-native speech. Non-native WER was reduced by half, while not affecting native performance Pronunciation Evaluation Combine scores using a regression tree to generate scores that correlate with scores from human raters Spectral Match: Compare the spectrum of a candidate phone to a native, context-independent phone model. Also used for mispronunciation detection Phone Duration: Compare the candidate duration to a model of native duration, normalized by rate of speech Speaking rate: phones/sentence 10/04/2006 41 Automatic Localization and Diagnosis of Pronunciation Errors Daniel Herron1, Wolfgang Menzel1, Erica Atwell2, Roberto Bisiani6, Fabio Deaneluzzi4, Rachel Morton5, Juergen Schmidt3, 1999 1U. of Hamburg, 2U. of Leeds, 3Ernst Klett Verlag, 4Dida*El S.r.l., 5Entropic, 6U. of Milan-Bicocca Locating and describing errors is critical for instruction Identifying segmental errors In response to a read prompt, lax recognition followed by strict recognition Some errors are predictable based on L1. Vowel, pre-vocalic consonant, and word-final devoicing errors are modelled explicitly, and tested on artificial data. Vowel - /ih/ -> /ey/ “it” -> “eet” PV consonant - /w/ - > /v/ “was” -> “vas” WF devoicing - /g/ -> /k/ “thinking” -> “thinkink” Using a word-internal tri-phone model vowel and PV consonant errors can be diagnosed with >80% accuracy with a FA rate less than 5%. WF devoicing can only be diagnosed with ~40% accuracy. Stress-errors are detected by deviation from trained models of stressed and unstressed syllables 10/04/2006 42 Automatic Syllable Stress Detection Using Prosodic Features for Pronunciation Evaluation of Language Learners Joseph Tepperman and Shrikanth Narayanan, 2005 University of Southern California Lexical stress can change POS and meaning “INsult” v. “inSULT” or “CONtent” v. “conTENT” Detecting stress on read speech with content determined a priori 10/04/2006 Use forced alignment to id syllable nuclei (vowels) Extract f0 and energy features. Duration features are manually normalized by context. Classified using a supervised Gaussian Mixture Model Post-processed to guarantee exactly 1 stressed classification per word. Mean f0, energy and duration discriminate with >80% accuracy on English spoken by Italian and German speakers 43 Nativeness, Fluency and Intonation Evaluation Intonational information can influence the proficiency and understandability of a secondlanguage speaker Proficient second-language speakers often have difficulty producing native-like intonation 10/04/2006 44 A Visual Display for the Teaching of Intonation Gerard Spaai and Dik Hermes, 1993 Institute for Perception Research Tools for guided instruction of intonation Intonation is difficult to learn It is acquired early, so it is resistant to change Native language intonation expectations may impair the perceptions of foreign intonation Intonation Meter Display the pitch contour. Interpolate non-voiced region Mark vowel onsets 10/04/2006 45 Quantitative Assessment of Second Language Learners' Fluency: An Automatic Approach Catia Cucchiarini, Helmer Strik and Lou Boves, 2002 University of Nijmegen Does ‘fluency’ always mean the same thing? Linguistic knowledge, segmental pronunciation, native-like intonation. Three groups of raters, 3 phoneticians, 6 speech therapists, assessed fluency of read speech on a scale from 1-10. With 1 exception the raters agreed with > .9 Native speakers are consistently rated as more fluent than non-native Time/Lexical correlates to fluency 10/04/2006 High rate of speech (segments/duration) High phonation/time (+/- pauses) ratio High mean length of runs Low number & duration of pauses 46 Prosodic Features for Automatic Text-Independent Evaluation of Degree of Nativeness for Language Learners Carlos Teixeira1,2, Horacio Franco2, Elizabeth Shriberg2, Kristin Precoda2, Kemal Sönmez2, 2000 1IST-UTL/INESC, 2SRI Can a model be trained to assess speakers nativeness similarly to humans without text information? Construct Feature-Specific Decision Trees Word stress (duration of longest vowel, duration of lexically stressed vowel, duration of vowel with max f0) Speaking rate approximations (durations between vowels) Pitch (max, slope “bigram” modeling) Forced alignment + pitch (duration between max f0 to longest vowel nucleus, location of max f0) Unique events (durations of longest pauses, longest words) Combination (max or expectation) of “posterior probabilities” from decision trees Results Pitch-based features do not generate human-like scores Inclusion of posterior recognition scores and rate of speech helps considerably. 10/04/2006 Only weak correlation (<.434) between machine and human scores Correlation = ~.7 47 Modeling and Automatic Detection of English Sentence Stress for Computer Assisted English Prosody Learning System Kazunori Imoto, Yasushi Tsubota, Antione Rau, Tatsuya Kawahara and Masatake Dantsuji, 2002 Kyoto University L1 specific errors need to be accounted for Japanese speakers tend not to use energy and duration to indicate stress Syllable structure “strike” -> /s-u-t-o-r-ay-k-u/ Incorrect phrasing Classification of stress levels Syllable alignment was performed with a recognizer trained with common native Japanese English speech (including segmental errors) Supervised HMM training using pitch, power, 4-th order MFCC & first and second order differences Using distinct models for each stress type/syllable structure/position combination (144 HMMs), 93.7%/79.3% native/non-native accuracies were achieved Two stage recognition increased accuracy to 95.1%/84.1% 10/04/2006 Primary + Secondary stress v. Non-stressed Primary v. Secondary stress 48 A study of sentence stress production in Mandarin Speakers of American English Yang Chen1, Michael Robb2, Harvey Gilbert2 and Jay Lerman2, 2001 1University of Wyoming, 2University of Connecticut Do native Mandarin speakers produce American English pitch accents “natively”? Experiment Compare native Mandarin English and native American English productions of “I bought a cat there” with varied location of pitch accent. Pitch Energy and duration of vowels were calculated and compared across language group and gender Vowel onset/offset were determined manually. Results 10/04/2006 Mandarin speakers produced stressed words with shorter duration than American speakers. Female mandarin speakers produced stressed words with greater rise in f0 49 Second Language Learning Systems Summary Performance assessment Pronunciation Intonation Error diagnosis and (Instruction) Influence of L1 on L2 instruction and evaluation 10/04/2006 50 Speech-to-Speech Translation ASR, MT and TTS components all exist independently Challenges specific to translation of speech Can speech information be used to reduce the impact of ASR errors on MT? Can information conveyed by intonation be translated via this framework? 10/04/2006 51 Speech-to-Speech Translation Paper List Cascaded Approaches Janus-III: Speech-to-Speech Translation in Multiple Languages A Unified Approach in Speech Translation: Integrating Features of Speech Recognition and Machine Translation Zhang et al. 2004 Explicit Use of Prosodic Information On the Use of Prosody in a Speech-to-Speech Translator Strom et al. 1997 A Japanese-to-English Speech Translation System: ATR-MATRIX Lavie et al. 1997 Takezawa et al. 1998 Integrated Approaches Finite State Speech-to-Speech Translation On the Integration of Speech Recognition and Statistical Machine Translation Matusov 2005 Coupling vs. Unifying: Modeling Techniques for Speech-to-Speech Translation 10/04/2006 Vidal 1997 Gao 2003 52 Cascaded Approach to Speech-to-Speech Translation ASR 10/04/2006 MT TTS 53 Janus-III: Speech-to-Speech Translation in Multiple Languages Alon Lavie, Alex Waibel, Lori Levin, Michael Finke, Donna Gates, Marsal Galvadà, Torsten Zeppenfeld , Puming Zhan, 1998 Carnegie Mellon University, University of Karlsruhe Interlingua and Frame-Slot based Spanish-English translation limited domain (conference registration) spontaneous speech Two semantic parse techniques GLR* Interlingua parsing (transcript 82.9%; ASR 54%) Phoenix (transcript 76.3%; ASR 48.6%) Identifies key concepts and their structure Parsing grammar contains specific patterns which represent domain concepts and a generation structure Phoenix is used as a back-off when GLR* fails. Manually constructed, robust grammar to parse input into interlingua Search for the maximal subset covered by the grammar Transcript: 83.3%; ASR 63.6% Late stage disambiguation 10/04/2006 Multiple translations are processed through the whole system. Translation hypothesis selection occurs just before generation using scores from recognition, parsing and discourse processing. 54 A Unified Approach in Speech-to-Speech Translation: Integrating Features of Speech Recognition and Machine Translation Ruiqiang Zhang, Genichiro Kikui, Hirofumi Yamamoto, Taro Watanabe, Frank Soong, Wai Kit Lo, 2004 ATR Process many hypotheses, then select one. In a cascaded architecture: Rescore MT hypotheses based on weighted log-linear combination of ASR and MT model scores HMM-based ASR produces N-best recognition hypotheses IBM Model 4 MT (a noisy channel model) processes all N. Construct the feature weight model by optimizing for a translation distance metric (mWER, mPER, BLEU, NIST) using Powell’s search algorithm Experiment Results Corpus: 162k/510/508 Japanese-English parallel sentences Baseline: no optimization of MT features Significant improvement was obtained by optimizing MT feature weights based on distance metric 10/04/2006 Additional improvement is achieved by including ASR features 55 Explicit Use of Prosodic Information How can prosodic information improve translation? How can prosodic information be translated? 10/04/2006 56 On The Use of Prosody in a Speech-to-Speech Translator Volker Strom1, Anja Elsner1, Wolfgang Hess1, Walter Kasper4, Alexandra Klein2, Hans Ulrich Krieger4, Jörg Spilker3, Hans Weber3 and Günther Görz3, 1997 1University of Bonn, 2University of Wien, 3University of Erlangen-Nürnberg, 4DFKI GmbH INTARC - German-English Translator produced for VERBMOBIL project. Spontaneous, limited domain (appointment scheduling) 80 minutes of prosodically labeled speech Phrase Boundary (PB) Detector Focus Detector Gaussian classifier based on F0, energy and time features with a 4 syl. window (acc. 80.76%) Rule based approach: Identifies location of steepest F0 decline (acc. 78.5%) Syntactic parsing search space is reduced by 65% 10/04/2006 Baseline syntactic parsing uses Decoder factor: product of acoustic and bi-gram scores Grammar factor: grammar model probability of a parse using the hypothesized word Prosody factor: 4-gram model of words and phrase boundaries 57 On The Use of Prosody in a Speech to Speech Translator Semantic parsing search space is reduced by 24.7% The semantic grammar was augmented, labeling rules as “segmentconnecting”(SC) and “segment-internal” (SI) SC rules are applied when there is a PB between segments, SI are applied when there are not. Ideal phrase boundaries reduced the number of hypotheses by 65.4% (analysis trees by 41.9%) Automatically hypothesized PBs required a backoff mechanism to handle errors and PBs that are not aligned with grammatical phrase boundaries. Prosodically driven translation is used when deep transfer (translation) fails A focused word determines (probabilistically) a dialog act which is translated based on available information from the word chain. 10/04/2006 Correct: 50%, Incomplete: 45%, Incorrect: 5% 58 A Japanese-to-English Speech Translation System: ATR-MATRIX Toshiyuki Takezawa, Tsuyoshi Morimoto, Yoshinori Sagisaka, Nick Campbell, Hitoshi Iida, Fumiaki Sugaya, Aiko Yokoo and Seiichi Yamamoto, 1998 ATR Limited domain translation system (Hotel Reservations) Cascaded approach ASR: sequential model ~2k word vocabulary MT: syntactically driven ~12k word vocabulary TTS: CHATR (concatenative synthesis) Early Example of “Interactive” Speech-to-Speech Translation. Speech Information is used in three ways in ATR-MATRIX Voice Selection Hypothesized phrase boundaries Using pause information along with POS N-gram information the source utterance is divided into “meaningful chunks” for translation. Phrase Final Behavior 10/04/2006 Based on the source voice, either a male or female voice is used for synthesis If phrase final rise is detected, it is passed to the MT module as a “lexical” item potentially indicating a question. 59 Integrated Approach to Speech-to-Speech Translation ASR+MT 10/04/2006 TTS 60 Finite-State Speech-to-Speech Translation Enrique Vidal, 1997 Universidad Politécnica de Valencia FSTs can naturally be applied to translation. FSTs for statistical MT can be learned from parallel corpora. (OSTIA) Speech input is handled in two ways: Baseline cascaded approach Integrated approach 1. Create an translation FST on parallel text 2. Replace each edge with an acoustic model of the source lexical item A major drawback of using this approach is large training data requirement. 10/04/2006 Align the source and target utterances, reducing their “asynchronicity” Cluster lexical items, reducing the vocabulary size 61 Finite-State Speech-to-Speech Translation Experiments Proof of concept experiment Text: ~30 lexical items used in 16k paired sentences (Spanish- English) Greater than 99% translation accuracy is achieved Speech: 50k/400 (training/testing) paired utterances, spoken by 4 speakers Best performance: 97.2% translation acc. 97.4% recognition accuracy Requires inclusion of source and target 4-gram LMs in FST training. Travel domain experiment Text: ~600 lexical items in 169k/2k paired sentences Speech: 336 test utterances (~3k words) spoken by 4 speakers 10/04/2006 0.7% translation WER w/ categorization; 13.3% WER w/o Text transducer was used, edges replaced by concatenation of “phonetic elements” modeled by a continuous HMM. 1.9% translation WER and 2.2% recognition WER were obtained. 62 On the Integration of Speech Recognition and Statistical Machine Translation E. Matusov, S. Kanthak and H. Ney 2005 Use word lattices weighted by HMM ASR scores as input to a weighted FST for translation Noisy Channel Model from source signal to target text TextTarget = argmax Pr(TextSource, TextTarget| Align) Pr(Signal| TextSource) Material: 4 parallel corpora Spontaneous speech in the travel domain 3k - 66k paired sentences in Italian-English, Spanish-English and Spanish-Catalan Vocabulary size 1.7k-15k words Results On all metrics (mWER, mPER, BLEU, NIST), the translation results are as follows: 1. 2. 3. 4. 5. 10/04/2006 Correct text Word lattice w/ acoustic scores Fully integrated ASR and MT (FUB Italian-English only) Word lattice w/o acoustic scores Single best ASR hypothesis (lower mPER than lattice w/o scores on FUB I-E) 63 Coupling vs. Unifying: Modeling Techniques for Speech-to-Speech Translation Yuqing Gao 2003 Application of discriminative modeling to ASR, with the goal of recognizing interlingua text for MT. Composing models (e.g., noisy channel models) can lead to local or suboptimal solutions Discriminative Modeling tries to avoid these by creating a single maximum entropy model p(text|acoustics,...) Includes other non-independent observations as features. Major considerations: 10/04/2006 To simplify computational complexity, acoustic features are quantized. Since the feature vector can get very large, reliable feature selection is necessary. In preliminary experiments, 150M features were reduced to 500K via feature selection 64 Speech-to-Speech Translation Summary Existing systems can be used to construct speech-tospeech translation systems However, two significant problems are encountered Intonational Information is generally ignored Prosodic Boundaries, Pitch Accent, Affect, etc. are important information carriers which ASR transcripts do not encode Local Minima The best recognized string may not generate the best translated string 10/04/2006 65 Intonation and Multi-Language Scenarios Use and Meaning of Intonation What information can intonation provide? Automatic Analysis of Intonation How can this information be represented computationally? Multi-Language Scenarios Second Language Learning Systems How can computers help teach a second language? Speech-to-Speech Translation How can machines translate speech? 10/04/2006 66 Thank you Questions. Automatic Analysis of Intonation Supervised? Detected Events Algorithm Yes Accent, Boundary DTree->HMM Ananthakrishnan Yes Accent, Boundary CHMM Ishi Yes 11 phrase final types DTree Shriberg Yes* Accent, Boundary, other. Many Taylor Yes* “Intonational Events” HMM Levow No Accent Spectral Clustering / Laplacian SVM Tamburini No Accent, Lex. Stress Threshold tuning Heldner Yes Pitch Accent Manual Rule Batliner Yes Accent, Boundary Decision Tree Wightman 10/04/2006 68 Second Language Learning Human corr? L1 Infl. Seg. Stress Timing / Duration Supra. Franco Yes No Yes No Yes No Herron Artificial Errors German / Italian Yes Yes No No Tepperman No No No Yes No No Spaai No No No Implicit Implicit Yes Cucciarini Yes No No No Yes No Teixeira Yes No Yes No Yes Yes Imoto No Japanese No Yes No Yes Chen No Mandarin No Yes Yes 10/04/2006 No 69 Speech-to-Speech Translation MT approach Cascaded / Integrated Languages Domain Lavie Interlingua Cascaded Japanese German Spanish Meeting Scheduling Zhang SMT Cascaded Japanese Travel Strom Interlingua Integrated German Meeting Scheduling Takezawa SMT Cascaded Japanese Hotel Desk Vidal SMT Integrated Spanish German Italian Travel Matusov SMT Integrated Italian Spanish Catalan Travel / Scheduling / Hotel Desk Gao Interlingua Generation Integrated NA NA 10/04/2006 70