Language
Phonemic Restoration Task
Also called: phoneme restoration, phonemic restoration effect, perceptual restoration of speech, added or replaced judgment, added-replaced discrimination
Measure how strongly listeners fill in a speech sound that noise covers: words with the sound under a burst of noise or replaced by it, told apart.
7-9 minUntimed
Free to start: 10 free participants per account, then $0.99 per participant for academic research, or any yearly plan. Try it here first, with no account.
- Papers, 2021 to 2025
- 19
- Evidence grade
- D
- Registered clinical studies
- 0
What it measures
Measures how strongly listeners fill in a sound that noise covers or replaces. Through headphones, a synthetic voice says everyday words in the participant's language with a burst of noise over one sound: the sound is either still there under the noise or taken out. Participants say which. The main score is how well they tell the two apart; lower means stronger restoration.
Constructs
- Perceptual restoration of speech sounds
- Use of word knowledge in listening
- Discrimination of speech sounds in noise
Research uses
- Speech perception research
- Reading and language research, at the group level
- Cross-language speech research
How the paradigm works
The Phonemic Restoration Task measures how strongly listeners fill in a speech sound that noise covers. When a sound is cut out of a spoken sentence and a cough put in its place, listeners still hear the missing sound: in Warren's (1970) study, 19 of 20 heard every sound present, and none could place the cough. Samuel (1981) made the illusion a measure. In some words a burst of noise is laid over a sound, which stays under it; in others the noise replaces the sound; listeners say which they heard. Someone who fills in the missing sound hears the two alike, so poor discrimination between them means strong restoration. It needs headphones and normal hearing. The test starts with the sound setup: the words load, the participant turns the device's volume until a steady tone is comfortable, and a six-trial headphone check (Woods, Siegel, Traer and McDermott, 2017) confirms that headphones are in use. By default a failed check ends the test, and its result is saved either way.
More about the paradigm
Each language has 36 everyday words of its own, written for the platform and said by that language's synthetic voice, not recorded from a person. Each word has one s sound inside it and no other hissing or rustling sound. The s was found by measurement: the one stretch of at least 30 ms where the word's energy from 3.5 to 8 kHz is within 15 dB of its highest and no more than 3 dB below its energy from 60 to 1,500 Hz. The noise covers that stretch with 15 ms more on each side, 61 to 255 ms across the languages, taking in the edges of the sounds next to it as Samuel's segments did. It is white noise at the s's own level, the match Del Tufo and Myers (2014) used. In the replaced version the noise takes the place of the s; in the added version it is laid over the s and the two together are turned down to the replaced version's level, so loudness does not tell them apart. Outside the noise, all three forms of a word are the same samples. Fricatives such as s are among the sounds that white noise restores most strongly (Samuel, 1981).
After the instructions, a demonstration plays one practice word as it is, with the noise added over the s, and with the s replaced. Eight practice trials with feedback follow, each of the four practice words in both versions. The test is two blocks of 32 words with a rest between them: each test word once in each block, its added version in one block and its replaced version in the other, so half the trials are each, in an order drawn from the session's seed with no more than three of one version in a row. Each trial plays its word once when the participant presses Play, and the answer, still there or replaced, has no time limit; nothing adapts to the answers. Every word plays at a fixed level, 12 dB below the volume-setting tone.
The main score is d prime: how well the replaced words were told from the added ones, the z of the share of replaced words answered replaced less the z of the share of added words answered replaced, both with the log-linear correction (Hautus, 1995). Lower means the two versions sounded more alike, stronger restoration; 0 is chance. The criterion beside it says which way the answers leaned: positive is a lean toward still there, toward hearing the words whole, which Samuel (1981) read as a decision made after hearing, apart from restoration itself. Discrimination of a fricative under white noise is low even in the laboratory, d prime 0.36 for fricatives in Samuel (1981) and 0.62 for words with a medial s in Del Tufo and Myers (2014), against 2.2 to 2.6 for the same noise segments heard alone. Many listeners score near chance here, so the measure is read across groups or conditions; an online version with ratings had a test-retest correlation of .37 (Giovannone and Theodore, 2023). Each language has its own words and voice, so results are compared within a language. The test does not measure or screen hearing: scores are research measures and do not diagnose.
Scores
- Discrimination (d prime)score
- z of the share of replaced words answered replaced less z of the share of added words answered replaced, both with the log-linear correction (Hautus, 1995).
- Lower means the two versions sounded more alike, stronger restoration; 0 is chance
- Criterionscore
- Minus the mean of the same two z values.
- Positive means a lean toward still there, toward hearing the words whole
- Replaced Answered Replaced%
- Share of the replaced words answered replaced.
- Informational
- Added Answered Replaced%
- Share of the added words answered replaced.
- Informational
- Percent Correct%
- Share of the answers that named the version played.
- 50 is chance
- Still There Answers%
- Share of the answers that were still there.
- Informational; half the words were each version
- Median Answer Timems
- Median time from when the answers opened, after the word, to the answer.
- Informational; the answer has no time limit
- Anticipationscount
- Answers within 100 ms of the answers opening, which leave the scores.
- Half or more of the answers makes the session invalid
- Correct Answerstrials
- Answers that count and named the version played.
- Significantly fewer than half (one-sided exact test) makes the session invalid
- Answers Countedtrials
- Test words answered 100 ms or more after the answers opened.
- Informational
- Completed Trialstrials
- Test words answered in the scored run.
- Fewer than planned is flagged for review
- Planned Trialstrials
- Test words the run set out to play.
- Informational
- Headphone Check Passed0/1
- 1 when the headphone check passed (at least 5 of 6 right), 0 when it did not.
- 0 stops the test unless the study set it to continue
- Headphone Check Correctcount
- Headphone check trials answered right, of 6.
- 5 or 6 passes
- Reported Output Delayms
- Output delay the device reported to the browser; empty when the browser does not report it.
- Informational
Use and evidence
19 papers from 2021 to 2025 name this task in their title or abstract, down 14% compared with 2016 to 2020. That ranks it 404 of 495 test families in our census. See the counts for every test.
Show the numbers
| Year | Papers |
|---|---|
| 2025 | 5 |
| 2024 | 2 |
| 2023 | 4 |
| 2022 | 6 |
| 2021 | 2 |
| 2020 | 2 |
| 2019 | 1 |
| 2018 | 6 |
| 2017 | 5 |
| 2016 | 8 |
| 2015 | 8 |
| 2014 | 5 |
| 2013 | 5 |
| 2012 | 3 |
| 2011 | 6 |
| 2010 | 2 |
| 2009 | 5 |
| 2008 | 3 |
| 2007 | 1 |
| 2006 | 4 |
| 2005 | 0 |
| 2004 | 2 |
| 2003 | 3 |
| 2002 | 1 |
| 2001 | 0 |
| 2000 | 0 |
Use grade E
Fewer than 30 papers from 2021 to 2025.
Evidence grade D
Little published measurement evidence found.
- Papers on test-retest reliability
- 0
- Papers on reliability or validity
- 2
- Registered clinical studies using it as an outcome
- 0
How the grades work
Use: A is 1,000 or more papers from 2021 to 2025, B 300 to 999, C 100 to 299, D 30 to 99 and E fewer than 30. Evidence: A means at least 25 papers report its test-retest reliability and at least 25 registered clinical studies use it as an outcome; B, at least 8 test-retest papers, and either 5 registered clinical studies or 150 reliability or validity papers; C, at least 2 test-retest papers, 30 reliability or validity papers or 3 registered clinical studies; D, little published measurement evidence found. The counts measure use. They include every paper that names the task in its title or abstract. Grades describe the published research on this task, not ConductCognition's version of it.
Source: ConductCognition test census 0.22.0, counted 3 November 2026 from OpenAlex and ClinicalTrials.gov.
Task design
The trial sequence as the test runs in ConductCognition.
- Sound setup
- The words load, the participant turns the device's volume until a steady 1 kHz tone is comfortable, then takes a 6-trial headphone check (Woods and colleagues, 2017). By default a failed check ends the test; the result is saved either way.
- Words
- 36 everyday words in each of the 8 languages, our own, each with one s sound inside it, said by that language's synthetic voice, not recorded from a person; 0.4 to 1.3 s each. 32 are for the test and 4 for the practice.
- Noise
- The s found by measurement, from its energy at 3.5 to 8 kHz, and widened by 15 ms on each side: 61 to 255 ms of white noise at the s's own level. Added: the noise over the s, the two together turned down to the replaced version's level. Replaced: the noise in place of the s. Outside the noise, all three forms of a word are the same.
- Trials
- Two blocks of 32 words with a rest between them: each test word once a block, its added version in one and its replaced version in the other, 16 of each a block, in a seeded order with at most 3 of one version in a row. Play, then still there or replaced, with no time limit.
- Practice
- A demonstration of one word as it is, added and replaced, then 8 trials with feedback for 2,000 ms.
- Level
- Every word 12 dB below the volume-setting tone, a level relative to the volume the participant set.
- Score
- d prime: the z of the replaced words answered replaced less the z of the added words answered replaced, log-linear corrected; lower is stronger restoration and 0 is chance. The criterion beside it; positive is a lean toward still there. Answers under 100 ms leave the scores.
The voices are synthetic and each language has its own words, so compare results within a language and within studies that use these words. White noise over an s is hard to tell from the s even in the laboratory, so many listeners score near chance: read the score across groups or conditions, not for one person. The level is relative to the volume the participant set. The test needs normal hearing and headphones, and it does not measure or screen hearing. The scores support research and do not diagnose.
Quality checks
Each session is marked valid, needs review or invalid, with the reason, before you see its scores.
Scored rows: phonemic_restoration_test from the scored run, one of two buttons after the word has played (still there or replaced), with no time limit; half the trials are each. Discrimination at chance is the strongest restoration, so accuracy is not flagged for being at chance; accuracy significantly below chance (one-sided exact test) invalidates, since it means the two answers were swapped. One answer on 90% of the trials is flagged, never invalid, since restoration leans listeners toward still there. Answers under 100 ms invalidate when they are at least half. A headphone check that stopped the test invalidates; a failed check with the test run anyway, or a run short of its planned trials, is flagged for review.
Why these rules
Restoration makes a replaced sound seem present, so a listener who restores strongly cannot tell the two versions apart, and chance accuracy is a result here (Samuel, 1981, found fricatives in white noise near it), while accuracy well below chance cannot come from restoration. Samuel also found restoration leaning listeners toward reporting the words whole. Accuracy significantly below chance (one-sided exact test, p < .05) cannot come from guessing; it means the response mapping was reversed or the task was not followed (Slick, Sherman and Iverson 1999 treat below-chance forced-choice performance as definite response bias). The answers open once the word has played and have no time limit, so an answer under 100 ms was pressed before they opened, and only a session made mostly of such answers is invalid. Woods, Siegel, Traer and McDermott (2017) let only listeners who passed the headphone check go on, because a listener on loudspeakers hears the sounds differently; a session that ran after a failed check keeps its scores but not its percentile.
No percentile. Scores and quality flags only.
Running this test
Settings you can change
| Setting | Standard | Allowed |
|---|---|---|
| When the headphone check fails | Stop the test | Stop the test, Continue the test |
Yearly plans can change settings within these limits. Free and pay-as-you-go studies use the standard settings.
Need a setting this test does not have?
Devices and languages
- Untimed test: runs on desktops, laptops, tablets and phones. Needs headphones. Participant devices
- Languages: English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese.
- Available on: Every plan, including the free plan.
- No percentile. Scores and quality flags only. How norms work
- Recruiting online? See the setup guides for Prolific, Sona Systems, CloudResearch Connect and Qualtrics.
Cite and describe
Participants completed the Phonemic Restoration Task in a web browser (ConductCognition task version 1.0.0, scoring version 1.0.0; Warren, 1970; Samuel, 1981), with the standard settings (when the headphone check fails: Stop the test). The main score was discrimination (d prime; z of the share of replaced words answered replaced less z of the share of added words answered replaced, both with the log-linear correction (Hautus, 1995)).
ConductCognition versions: task version 1.0.0, scoring version 1.0.0. Every export records the versions each session used, for a methods section.
- Warren, R. M. (1970). Perceptual restoration of missing speech sounds. Science, 167(3917), 392-393. https://doi.org/10.1126/science.167.3917.392The illusion: a sound replaced by a cough is still heard, and the cough cannot be placed.
- Samuel, A. G. (1981). Phonemic restoration: Insights from a new methodology. Journal of Experimental Psychology: General, 110(4), 474-494. https://doi.org/10.1037/0096-3445.110.4.474The paradigm: added or replaced judgments scored by signal detection; fricatives under white noise restored most.
- Del Tufo, S. N., & Myers, E. B. (2014). Phonemic restoration in developmental dyslexia. Frontiers in Neuroscience, 8, 134. https://doi.org/10.3389/fnins.2014.00134A medial s under white noise matched to it in level and length.
- Hautus, M. J. (1995). Corrections for extreme proportions and their biasing effects on estimated values of d'. Behavior Research Methods, Instruments, & Computers, 27(1), 46-51. https://doi.org/10.3758/BF03203619The log-linear correction of the two shares.
- Woods, K. J. P., Siegel, M. H., Traer, J., & McDermott, J. H. (2017). Headphone screening to facilitate web-based auditory experiments. Attention, Perception, & Psychophysics, 79(7), 2064-2072. https://doi.org/10.3758/s13414-017-1361-2The headphone check before the test.
- Giovannone, N., & Theodore, R. M. (2023). Do individual differences in lexical reliance reflect states or traits? Cognition, 232, 105320. https://doi.org/10.1016/j.cognition.2022.105320Phonemic restoration run online, with its test-retest reliability.
Source papers
- Samuel AG (1981). Phonemic restoration: Insights from a new methodology. Journal of Experimental Psychology: General, 110(4):474-494.N 20, Adults
- Del Tufo SN, Myers EB (2014). Phonemic restoration in developmental dyslexia. Frontiers in Neuroscience, 8:134.N 53, College age (mean 19.7)
Questions
How long does the Phonemic Restoration Task take?
About 7 to 9 minutes, as listed in the test library.
What device does a participant need?
Untimed test: runs on desktops, laptops, tablets and phones. Needs headphones. The browser must be Chrome, Edge, Firefox or Safari.
What scores does it produce?
Discrimination (d prime), Criterion, Replaced Answered Replaced, Added Answered Replaced, Percent Correct, Still There Answers, Median Answer Time, Anticipations, Correct Answers, Answers Counted, Completed Trials, Planned Trials, Headphone Check Passed, Headphone Check Correct and Reported Output Delay. Each session is also marked valid, needs review or invalid, with the reason.
How widely is the Phonemic Restoration Task used in research?
19 papers from 2021 to 2025 name it in their title or abstract, down 14% compared with 2016 to 2020. That ranks it 404 of 495 test families in the ConductCognition test census (use grade E). Source: ConductCognition test census 0.22.0, counted 3 November 2026 from OpenAlex and ClinicalTrials.gov.
How much reliability evidence does the Phonemic Restoration Task have?
In the published research, 0 papers report its test-retest reliability and 0 registered clinical studies use it as an outcome (evidence grade D in the ConductCognition test census).
Does it report a percentile?
No. Scores and quality flags only.
Which settings can a researcher change?
When the headphone check fails (Stop the test, Continue the test). Settings can be changed on yearly plans; free and pay-as-you-go studies use the standard settings.
Which languages can participants use?
Participants can take this test in English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese, with its words in the same language. English is on every plan, and the other languages come with yearly plans. Have a native speaker read the machine-translated text before launch.
Is this a clinical test?
No. ConductCognition is for research use only. It runs, scores and exports the test; clinical interpretation is outside the platform.
Every test: the full test list.
Run this test in your next study
Start free with 10 free participants on any test, then use participant credits or a yearly plan.