Speech Intelligibility and Repair Assessment for Voice-Process, Support and Client-Facing RolesThe listener never lost your accent. They lost the word. This assessment finds out whether you can hear which word, name why, and fix it.

Methodology: Forty-four original items across seven formats: eighteen detection trials on a fixed closed option set, six select-every-change repair items, five match-the-following, five true-or-false statements of current guidance, four ordering items, three listening items with a written equivalent, and three spoken items graded on the transcript and reported apart. CONSTRUCT STATEMENT: this measures whether a speaker can tell what a listener will lose in connected speech, name the mechanism that costs the listener the word, and change it so the listener does not lose it; it does not measure accent, nationality, vocabulary range, grammar, or how the reader personally sounds. DECLARED RESPONSE INSTRUCTION: knowledge, one instruction for the whole instrument. Every keyed item asks what a listener will lose, which mechanism causes it, or which change removes it, judged against the line as described; no item asks the reader to rate themselves and none is a behavioural-tendency item. FRAMEWORK ANCHOR: the Common European Framework of Reference for Languages, Companion Volume (Council of Europe, 2020), scale for phonological control, whose three sub-scales are overall phonological control, sound articulation and prosodic features, and whose organising criterion is intelligibility rather than nativeness. The five reported mechanisms are placed on it: speaking rate and pausing, word stress and thought grouping under prosodic features; consonant endings and clusters and vowel contrast under sound articulation. CEFR is cited as a published framework, with no implication of endorsement by the Council of Europe. No commercial English test is named, reproduced or referred to anywhere in this instrument. SCORING DESIGN: C6 signal detection with a NAMED hit condition, on the eighteen detection trials. Each trial prints a line as it was said, with three features of its delivery, and the same six categories as options on every trial: the five mechanisms and 'nothing is lost'. Twelve trials carry one real mechanism among two harmless accent features (signal); six carry only harmless features (noise). A HIT is a signal trial on which the reader flagged the line AND named the mechanism that costs the listener the word; a signal trial on which the reader named a different mechanism is counted as SENSED BUT MISNAMED and is a miss for d-prime; a noise trial on which the reader named any mechanism is a FALSE ALARM; naming nothing on a noise trial is a CORRECT REJECTION. d' = z(hit rate) - z(false alarm rate); c = -0.5 (z(hit rate) + z(false alarm rate)); the log-linear correction adds 0.5 to every cell before a rate is taken, so an extreme rate cannot make z infinite. Both are reported, because a flagger and a filterer can share a d-prime. THE NAMING REQUIREMENT AND THE ORDINARY RESPONDENT (playbook §6.32, §6.79): requiring the mechanism to be named stops indiscriminate flagging from earning a high hit rate, and it also puts the prior-drawing respondent below zero on raw d-prime. The declared marginals put that respondent at d' of about -0.46 on a full sitting, and the say-nothing habit at about -0.30, so the two are within two tenths of each other; naming a fixed mechanism on every trial lands between -2.0 and -2.4 and naming at random lands near -1.7. The headline is therefore d-prime AGAINST THE ORDINARY RESPONDENT: the reader's d-prime minus the d-prime of a respondent drawing every answer from the declared per-option marginals on the same answered trials, with the same correction. Zero on that scale is the ordinary respondent, not the coin. Raw d-prime and c are printed beside it. REPAIR STRAND, reported apart: six select-every-change items, four ordering items and one match-the-following, each reduced to a quality in [0,1] over the answered item (the share of lines whose ticked state matches the key; the share of positions correct; the share of pairs correct) against the expected quality of a respondent drawing from the item's declared per-line, per-position or per-pair marginal; chance-corrected as 100 (mean quality - mean chance) / (1 - mean chance), with its own omega and never averaged into d-prime. THE REGISTER: each mechanism pools every keyed unit tagged to it (its signal trials, the noise trials whose lure it is, its repair items, its guidance statement, its listening item, and the left-hand items of a matching exercise that name it) into a chance-corrected figure with a declared modelled standard deviation, an assumed omega and a three-way placement; a mechanism with fewer than eight answered units, or omega under .70 on the unrounded value, carries a placement word and no figure. A mechanism with no answered unit at all is printed as not evidenced today, full size, in its own row. Rows are ordered by the mechanism's DECLARED cost to a listener, printed above the register with its sources, never by the reader's score. THE SPOKEN ROUND: three voice_response items, transcribed and rubric-graded on the transcript against four content criteria each. The platform places no pronunciation, accent, pace or fluency score on a daily-catalogue voice response, and the rubric asks only about what the transcript can carry: whether the line was said in full, where the pauses were named, which mechanism was named, which change was named, and that the aim was not described as sounding native. The spoken round is reported entirely apart, enters no keyed figure, and prints DEFERRED in plain words where the transcription or the grading did not run; a zero with no transcript is nobody listening, never a poor answer. THE NULL: every chance correction runs against the item's declared marginal: option_base_rate on the detection trials, the guidance statements and the listening items, select_prior per line, pair_base_rate per left-hand item, position_base_rate per position. These are authored assumptions stated in the open; observed shares replace them once live data exists. RELIABILITY: McDonald's omega in its Spearman-Brown form from the answered unit count and an assumed average inter-item correlation of .24, printed as an assumption; the standard error of d-prime is the larger of its binomial standard error (Gourevitch and Galanter 1967) and the omega-based one on a declared modelled standard deviation. Every figure carries a 68 and a 95 per cent band, drawn and written. REFUSALS: an unanswered item leaves numerator, denominator and chance term; an empty sitting scores exactly zero everywhere and is not reportable; fewer than twenty-two answered keyed items is withheld; the detection figures need at least ten answered trials with at least six signal and four noise, else a placement word; no percentile appears anywhere because there is no norm group; no repair keyed anywhere amounts to sounding less Indian, because the construct is intelligibility and CEFR's whole point is that intelligibility is not nativeness. LOCALE: every line is one an agent on an Indian support floor would say to a listener in the UK, the US or Australia; the harmless features on the noise trials (retroflex stops, unaspirated stops, a syllable-timed rhythm at a moderate pace, a sounded r, th as t in 'thank you', a short a in 'can't', the British stress on 'advertisement') are accent features attested in the descriptive literature that a listener in those countries follows without effort. SOURCES: Council of Europe, Common European Framework of Reference for Languages: Learning, teaching, assessment, Companion Volume (2020), phonological control scale and its three sub-scales; Jenkins, The Phonology of English as an International Language (Oxford, 2000), the Lingua Franca Core, for consonant clusters, vowel length and nuclear stress as the segmental and prosodic features that carry intelligibility; Derwing and Munro, Pronunciation Fundamentals: Evidence-based Perspectives for L2 Teaching and Research (John Benjamins, 2015), and Munro and Derwing, Foreign accent, comprehensibility and intelligibility in the speech of second language learners (Language Learning, 1995), for the separation of accentedness from intelligibility; Munro and Derwing, Modeling perceptions of the accentedness and comprehensibility of L2 speech: the role of speaking rate (Studies in Second Language Acquisition, 2001), for rate; Derwing, Munro and Wiebe, Evidence in favor of a broad framework for pronunciation instruction (Language Learning, 1998), for prosody-focused instruction outperforming segmental instruction on comprehensibility; Hahn, Primary stress and intelligibility: research to motivate the teaching of suprasegmentals (TESOL Quarterly, 2004); Field, Intelligibility and the listener: the role of lexical stress (TESOL Quarterly, 2005); Cutler, Native Listening: Language Experience and the Recognition of Spoken Words (MIT Press, 2012), for stress-based lexical access; Brown, Functional load and the teaching of pronunciation (TESOL Quarterly, 1988), and Munro and Derwing, The functional load principle in ESL pronunciation instruction (System, 2006), for which vowel contrasts cost; Wells, Accents of English (Cambridge, 1982), volume 3, for the descriptive account of Indian English features treated as harmless here, and Wells, Longman Pronunciation Dictionary (3rd edition, 2008), for the British and American stress variants; Gourevitch and Galanter, A significance test for one parameter isosensitivity functions (Psychometrika, 1967), for the standard error of d-prime; Hautus, Corrections for extreme proportions and their biasing effects on estimated values of d' (Behavior Research Methods, 1995), for the log-linear correction; Macmillan and Creelman, Detection Theory: A User's Guide (2nd edition, 2005). All items are original works written for this instrument. No trademarked instrument, test or scale is reproduced or named; the features, lines and repairs are the authors' own. Not affiliated with, endorsed by or derived from any commercial English test or any certification body.