2 Topics Original FDA overview Advantages and disadvantages of this assessmentFDA 2 – new aspectsComputerised FDADemonstrationCurrent work on automated intelligibility testing
3 Original FDA Author: Pam Enderby First published in 1983 Result of research identifying nature and patterns of oromotor movements associated with different neurological diseases (Enderby 1983)Translated into French, German, Dutch, Norwegian, Swedish, Finnish, Catalan and Castilian
4 Aim of FDATo analyse several important parameters of the motor speech systemTo guide treatmentTo assist with neurological diagnosisTo have good reliability and validity between and within clinicians without extensive training
5 Structure of FDA Reflexes Respiration Lips Palate Laryngeal Tongue Cough, swallow, dribble/droolRespirationAt rest, in speechLipsAt rest, spread, seal, alternate, in speechPalateFluids, maintenance, in speechLaryngealTime, pitch, volume, in speechTongueAt rest, protrusion, elevation, lateral, alternate, in speechIntelligibilityWords, sentences, conversation
6 Procedure Ask patient to carry out a task Rate ability of each parameter using a 9 point scale – 5 descriptors + ½ marks
7 Advantages of FDAIntelligibility commonly used to assess severity of dysarthria and to monitor progress BUT Intelligibility measures alone do not diagnose type of dysarthria or guide treatmentFDA breaks speech up into its component parts so the clinician can analyse what contributes to the reduced intelligibility thus guiding treatmentFDA provides a profile that contributes to the neurological diagnosis
8 Disadvantages of FDA Some measures can be subjective Some descriptors are interpreted differently by different clinicians reducing reliabilityIntelligibility section:Too few words/sentences regular users can learn themSentence structure = ‘the man is…’ therefore only listening for the last wordScoring system based on number listener understood out of 10 (crude)
9 FDA 2 Authors: Pam Enderby & Rebecca Palmer 2008 Aim: To address theoretical and practical issues identified in reviews of the first edition
10 Improvements 1Omitted items that have been found to be unreliable or redundant to the purposes of diagnosis and treatmente.g. Jaw tests – patients rarely have abnormality in the jaw therefore the information didn’t assist diagnosis
11 Improvements 2 Improved reliability of descriptors Inter-rater reliability testing between experienced users of the FDA showed that some descriptors were interpreted differently.E.g. voice timePatient can say ‘ah’ for 15 secondsPatient unable to sustain clear voice for 3 secondsConstant hoarse voice – RP = a), PE = e)
12 Improvements 2 Inter rater and test retest reliability Audio recordings of 9 people with a range of types and severities of dysarthria performing the audible FDA 2 tests:6 speech therapists working with a mixed adult caseload judged 42 examples of FDA 2 tests.Scored on a 9 point scaleSame 42 tests presented again to the listeners after 6 week intervalInter and intra rater reliability were calculated using intra class correlation coefficients
13 Inter and intra judge reliability 1234560.760.770.920.560.650.720.670.600.51-0.380.520.490.790.730.660.70Criteria for interpretation of reliability coefficients for ordinal measures (Landis & Koch, 1977):<0 = poor, = slight, = fair,= moderate (mod), = substantial (sub)0.81 – 1 = almost perfect (per)
14 Improvements 3 In speech tests Sound saturated sentences provided for patient to say so that clinician can listen to the accuracy of sound placement in speechLips in speech:‘Mary brought me a piece of maple syrup pie’Tongue in speech:‘Kenneth’s dog took ten tiny ducks today’
15 Improvements 4 Intelligibility testing New set of words Corpus of 116 words to reduce probability of listeners learning the words with increased exposurePhonetically balanced list for types of sounds, position of sounds in words, word lengthWord frequency >10 per million to control for any effects of word frequency on intelligibility
16 Improvements 4 Sentence intelligibility Key words phonetically balanced to account for place, manner, position and word lengthCarrier phrases/sentences are all different so the listener has to listen to a sentence, not just interpret the key word in a standard carrier phrase‘Can you go the shop?’‘My daughter is a nurse’‘Lets go to the theatre’
17 AvailabilityFDA 2 available now from Pro-edOnly in English!
18 Computerised FDA James Carmichael produced computer version Demonstration
19 Planned additions to CFDA Automation of intelligibility testing – modelling the naiive listener If the learning effect alters a listener’s perception of a particular individual’s speaking style, is that listener’s judgement still representative of the naïve listener?Can a computer model be built which behaves like an “eternal” naïve listener (i.e. never adapting to an unfamiliar speaking style and therefore always consistent in assessment)?
20 Using HMM Models to Emulate the Naïve listener A hidden Markov Model (HMM)a statistical representation of a speech unit at the phone/word/utterance level.HMM models are “trained” by analysing the acoustic features of multiple utterances representing the specified speech unit.'Everyman' HMM Word ModelMultiple Speech Samples from multiple speakers
21 Goodness of fitOnce trained, an HMM word model can be used to estimate the likelihood that a given speech sound could have actually been produced by that word model.This likelihood is called a goodness of fit (GOF)expressed as a log likelihood, e.g (or simply expressed as -35).
22 Comparing GOF scores with Subjective Assessments 3 important cues of intelligibility are:hesitation time;speech ratea phoneme-by-phoneme comparison of what the speaker intended to say and what the listener actually heard.
23 Calculating Phonetic Convergence Phoneme comparison of intended and perceived message: “You have to pay”(for a mildly dysarthric speaker)Intended/j//u://h//æ//v//t//p//e/Heard/d//b//aι/Convergence1Word Level Deletion-1Overall Convergence5 out of a possible 9 = 0.56 (56%)
24 Phonetic convergenceHesitationL1L5L10L15L20ListenersL1L5L10L20L15ListenersMild, Moderate, SevereMild, Moderate, SevereSpeech rateSpeech rate’s correlation with intelligibility is not as good as hesitation time or phonetic convergence, so we derive a Perceptual Intelligibility Index (PII) based on the Phonetic Convergence score weighted by a hesitation time coefficientMild, Moderate, Severe
25 How well do automated GOF scores correlate with Perceptual intelligibility index? SpeakerPhon. ConvergenceHesitation Time coefficientSentence PII ScoreAvg. GOF ScoreMild0.950.910.86-34Moderate0.270.150.11-61Severe0.200.190.04-85Correlation between GOF scores and PII scores =0.72Automated scores of goodness of fit measures generated by HMMs could be a valid and consistent intelligibility measure
26 Summary FDA 2 Analyses each parameter of speech Enables clinician to find cause of reduced intelligibility, guiding treatmentAssists with diagnosis of dysarthria type and neurological impairmentExcludes redundant testsUses non-ambiguous descriptorsHas inter and intra-rater reliabilityLarge corpus of words and sentences controlled for linguistic and phonetic parameters for intelligibility sectionsWord and sentence cards provided
27 Summary Computerised FDA Provides training test for new users Automatically produces profile and stores informationIncreases objectivity of measuresProvides visual feedback of performance and improvements to patientSeeks to automate measurement of intelligibility leading to increased consistency