Speech Intelligibility

Abstract

Speech intelligibility is the proportion of spoken material a listener correctly recognizes, a perceptual outcome distinct from mere audibility. Early work at Bell Laboratories reduced it to the audible fraction of the speech spectrum, yielding the articulation index and its modern successor, the Speech Intelligibility Index (French & Steinberg, 1947; Kryter, 1962). Later measures captured the room and the channel through the modulation transfer function and the Speech Transmission Index (Houtgast & Steeneken, 1985). Intelligibility is not fixed by the acoustic signal alone: context and word predictability shift it sharply (Miller, Heise, & Lichten, 1951), and recognition under degradation exacts a measurable cognitive cost (Peelle, 2018). This article surveys how intelligibility is measured, predicted, and applied across audiology, telecommunications, and clinical speech assessment.

Keywords: speech intelligibility, articulation index, speech reception threshold

When a talker and a listener share a quiet room, almost every word arrives intact. Add background babble, a reverberant hall, a telephone codec, or a motor-speech disorder, and recognition falls — but not uniformly, and not in step with loudness. A shout in noise can be less intelligible than a whisper in quiet. Speech intelligibility is the quantity that captures this: the fraction of linguistic units a listener actually identifies, measured behaviorally or predicted from the physics of the signal and the channel (French & Steinberg, 1947). It is one of the oldest quantitative problems in the psychology of hearing, and it remains central wherever speech must cross a degrading medium.

What Speech Intelligibility Is

Speech intelligibility is an operational measure of successful spoken communication: present a listener with speech material, count the proportion of units — phonemes, words, or key words in sentences — reported correctly, and that proportion is the intelligibility score under those conditions (French & Steinberg, 1947). The definition is deliberately behavioral. It makes no commitment to how the listener recovers the message; it records only whether the message arrived.

Three distinctions give the term its precision. First, intelligibility is not audibility. A sound can be comfortably audible yet unintelligible, because noise, distortion, or reverberation has smeared the cues that distinguish one phoneme from another. Second, intelligibility is not quality or naturalness; a synthetic voice may sound mechanical while remaining perfectly intelligible, and a warm, natural recording may be unintelligible in a bad room. Third, intelligibility is a property of the whole talker–signal–channel–listener system, not of the speech alone: the same recording yields different scores for different listeners, channels, and competing sounds.

Because it is behavioral, intelligibility is also statistical. A single score is a point on a psychometric function that rises with the signal-to-noise ratio, and the full function — not any one number — characterizes a condition. Much of the measurement literature is concerned with how to summarize that function economically and how to predict it without running the full behavioral test.

Measuring Intelligibility

Direct measurement presents calibrated material and scores the response. The material determines what the score means. Nonsense syllables and phonetically balanced word lists strip away context, so they probe the acoustic–phonetic channel itself and are sensitive to fine distortions. Everyday sentences restore the redundancy of real language, so they score higher at the same signal-to-noise ratio and better predict functional communication. The gap between the two is not noise; it is the contribution of linguistic knowledge, and it is large and lawful (Miller, Heise, & Lichten, 1951).

A landmark tool exploits that gap directly. The Speech Perception in Noise test pairs sentences whose final word is highly predictable from context with sentences whose final word is unpredictable, holding the target words otherwise comparable (Kalikow, Stevens, & Elliott, 1977). The difference between the two scores isolates a listener's ability to use context — a measure that distinguishes, for instance, listeners whose audiograms look alike but who differ in real-world difficulty.

Both conditions follow the same S-shaped law, but context slides the whole function leftward: at the same signal-to-noise ratio the high-context score is far higher, because the listener predicts what the acoustics leave ambiguous. The horizontal distance between the curves at 50% is the context benefit in decibels.

Running a full psychometric function is costly, so clinical practice summarizes it with a single adaptive number: the speech reception threshold, the signal level or signal-to-noise ratio yielding 50% correct. An adaptive staircase makes the target harder after a correct response and easier after an error, so the presentation level hovers around the threshold and converges on it in relatively few trials. Averaging the reversal levels estimates the threshold efficiently and reliably, and standardizing the sentence materials sharpened that reliability further (Plomp & Mimpen, 1979). The threshold's appeal is that one interpretable number, such as a listener needing the speech 3 dB louder than the noise to get half the sentences, replaces a whole curve.

Predicting Intelligibility from the Signal

A century of engineering effort has gone into predicting intelligibility without a listener, from measurable properties of the signal and channel. The founding idea, developed at Bell Laboratories, is that intelligibility depends on how much of the speech spectrum is both audible and undistorted (French & Steinberg, 1947). Divide the frequency range into bands, find in each the fraction of the speech's dynamic range that rises above the noise and above the listener's threshold, weight each band by how much it contributes to intelligibility, and sum. The result — the articulation index, a number between 0 and 1 — predicts the intelligibility score a listener would obtain. Kryter formalized the calculation into a standard procedure, fixing the band divisions and weights that made the index reproducible across laboratories (Kryter, 1962). Its modern standardized descendant is the Speech Intelligibility Index.

With all five bands clear the index is 0.64 — the worked-example value. Masking the noisiest bands drops it toward 0, and the two predicted scores fall apart: sentence intelligibility holds up far longer than nonsense-syllable intelligibility, because sentence redundancy compensates for the lost bands.

The articulation index captures additive noise and filtering, but it is blind to two things that wreck intelligibility in real rooms: reverberation and nonlinear distortion, which can leave every band nominally audible while smearing speech across time. The solution was to measure the channel differently. Speech carries its information in slow fluctuations of amplitude — the envelope — and a good channel preserves those fluctuations while a reverberant or distorting one flattens them. The modulation transfer function quantifies how much the channel reduces modulation depth across modulation frequencies, and the Speech Transmission Index condenses that function into a single 0-to-1 predictor of intelligibility in that room or over that link (Houtgast & Steeneken, 1985). It is now the standard for rating auditoria, public-address systems, and voice communication channels.

Context, Effort, and the Clinic

Signal-based measures treat the listener as a fixed decoder, but listeners are not fixed. The single most robust behavioral finding is that context and predictability raise intelligibility sharply: a word that can be guessed from its sentence is recognized at signal-to-noise ratios where the same word in isolation is lost (Miller, Heise, & Lichten, 1951). This top-down contribution is why sentence tests and isolated-word tests diverge, why the Speech Perception in Noise test works, and why any purely acoustic index is an incomplete predictor of real-world communication.

Recognition that succeeds can still be costly. Understanding speech in noise, through a hearing aid, or across a bad line recruits attention and working memory, and that recruitment can be measured — through pupillometry, dual-task performance, and self-report — even when the intelligibility score is at ceiling (Peelle, 2018). The Framework for Understanding Effortful Listening organized these findings, treating listening as the deployment of limited cognitive resources against acoustic challenge and linking hearing impairment to fatigue and disengagement (Pichora-Fuller et al., 2016). Two listeners with identical scores can differ sharply in the effort those scores cost.

Clinically, intelligibility is also an outcome of the talker, not only the listener. In the motor-speech disorders such as dysarthria, intelligibility quantifies how a disorder degrades communication and how therapy restores it. Systematic phonetic scoring — identifying which sound contrasts a disordered talker fails to convey — turns a global percentage into a diagnostic profile that points at the underlying impairment (Kent, Weismer, Kent, & Rosenbek, 1989). Here intelligibility is at once an outcome measure and a window on the speech-production system.

Each correct response (green) makes the next sentence harder by stepping the signal-to-noise ratio down; each error (red) steps it back up. The track soon oscillates around the true threshold, and averaging the reversal points — where the direction changed — estimates the speech reception threshold in far fewer trials than mapping the whole psychometric function would take.

Worked Example

Consider a simplified articulation-index calculation over five frequency bands of equal importance, each carrying weight 0.20. Take the long-term speech peak in every band as 65 dB and assume the audible speech cues span a 30 dB dynamic range below that peak (so speech information lives between 35 and 65 dB). In each band, the audible fraction is the portion of that 30 dB range standing above the noise floor.

Suppose the noise floors are 50, 41, 35, 56, and 47 dB in the five bands. The audible fraction in a band is (65 − noise floor) ÷ 30, clipped to the range 0 to 1:

- Band 1: (65 − 50) ÷ 30 = 15 ÷ 30 = 0.50 - Band 2: (65 − 41) ÷ 30 = 24 ÷ 30 = 0.80 - Band 3: (65 − 35) ÷ 30 = 30 ÷ 30 = 1.00 - Band 4: (65 − 56) ÷ 30 = 9 ÷ 30 = 0.30 - Band 5: (65 − 47) ÷ 30 = 18 ÷ 30 = 0.60

The articulation index is the weighted sum: 0.20 × (0.50 + 0.80 + 1.00 + 0.30 + 0.60) = 0.20 × 3.20 = 0.64. An index of 0.64 predicts high intelligibility for everyday sentences — well above 90% — but only moderate scores for nonsense syllables, because the redundancy of sentences compensates for the roughly one-third of the speech cues the noise has masked. Raising the level of the two noisiest bands, or lowering their noise, would lift the index toward 1.0 and close that gap; the single number tracks exactly how much usable speech information the channel delivers.

Discussion

The measurement of intelligibility has proceeded along two converging lines. One is behavioral: present material, count correct units, and summarize the resulting psychometric function — economically, through an adaptive threshold. The other is predictive: measure the signal and channel and compute an index that forecasts the behavioral score without a listener. The articulation index, the Speech Intelligibility Index, and the Speech Transmission Index are successive refinements of the predictive line, each capturing a degradation the previous one missed — additive noise, then filtering, then reverberation and nonlinear distortion.

The enduring tension is that prediction treats the listener as a fixed acoustic decoder, while the strongest behavioral result is that the listener is anything but fixed. Context, predictability, linguistic knowledge, attention, and working memory all move the score, sometimes by more than a large change in the signal would. The signal-based indices remain indispensable for engineering a room or a codec, because they isolate the part of intelligibility the designer controls. But they systematically under-describe real communication, where a knowledgeable, attentive listener extracts a message from acoustics that any index would score as marginal — at a cognitive cost the index cannot see.

Current Directions

The most active front reframes intelligibility as one output of a resource-limited cognitive system rather than a property of the signal. The listening-effort program treats accurate recognition and the effort it costs as separable outcomes, and asks when effort predicts fatigue, disengagement, and poorer real-world function better than the intelligibility score itself (Peelle, 2018). The Framework for Understanding Effortful Listening gave this work a common vocabulary, drawing hearing science together with the psychology of motivation and cognitive load (Pichora-Fuller et al., 2016).

A parallel direction is computational and clinical. Objective, instrumental predictors of intelligibility are increasingly built from machine-learning models trained on large speech corpora, aiming to match listener scores for hearing-aid processing and for synthetic and disordered speech. In the motor-speech clinic, automatic phonetic analysis extends the diagnostic scoring logic to continuous speech and larger caseloads (Kent, Weismer, Kent, & Rosenbek, 1989). The open problem common to both is the one this field has circled since the 1950s: building the listener's knowledge, context, and effort into a predictor that still runs without the listener.

Common Misconceptions

Louder speech is more intelligible speech.

Only up to a point, and only in quiet. In noise, what matters is the signal-to-noise ratio, not absolute level; raising both speech and noise together leaves intelligibility unchanged, and very high levels can reduce it. Audibility is necessary but not sufficient (French & Steinberg, 1947).

A good intelligibility score means communication was effortless.

Score and effort are separable. A listener can reach 100% correct while spending heavily on attention and working memory, a cost measurable by pupillometry and dual-task methods and invisible to the percentage score (Peelle, 2018).

An acoustic index like the STI fully predicts how well people will understand.

Signal-based indices predict the acoustic contribution well but omit context, linguistic knowledge, and listener state, which move real scores substantially (Miller, Heise, & Lichten, 1951). They describe the channel, not the whole communication.

Intelligibility is a property of the talker's speech.

It is a property of the whole talker–signal–channel–listener system. The same recording scores differently across listeners, rooms, and competing sounds, which is why a condition must be specified before a score means anything (Kent, Weismer, Kent, & Rosenbek, 1989).

Glossary

Articulation index (AI).

A predictive measure, between 0 and 1, formed by weighting the audible and undistorted fraction of the speech spectrum across frequency bands; the foundational signal-based predictor of intelligibility.

Audibility.

Whether a sound exceeds the listener's threshold and any masking noise. A precondition for intelligibility but not equivalent to it.

Babble.

Competing speech from several talkers, a common masking noise in intelligibility testing because it adds both energetic and informational masking.

Dynamic range.

The span in decibels between the peaks and troughs of the speech signal that carry usable information; roughly 30 dB in articulation-index models.

Dysarthria.

A motor-speech disorder in which impaired control of the speech musculature degrades intelligibility; assessed by phonetic scoring of which sound contrasts fail.

Intelligibility.

The proportion of presented speech units a listener correctly recognizes under specified conditions.

Listening effort.

The cognitive resources deployed to recognize speech under challenge, measurable even when the intelligibility score is at ceiling.

Modulation transfer function (MTF).

A description of how much a channel or room reduces the depth of amplitude modulations across modulation frequencies; the basis of the Speech Transmission Index.

Phonetically balanced list.

A word list whose phoneme frequencies approximate those of the language, used so that a score reflects the acoustic-phonetic channel rather than any single sound.

Predictability.

The extent to which context constrains an upcoming word; high predictability raises intelligibility at a given signal-to-noise ratio.

Psychometric function.

The S-shaped curve relating percent correct to a stimulus variable such as signal-to-noise ratio; the full characterization of an intelligibility condition.

Signal-to-noise ratio (SNR).

The level of the target speech relative to the masking noise, in decibels; the primary determinant of intelligibility in noise.

Speech Intelligibility Index (SII).

The modern standardized successor to the articulation index, computing the usable fraction of the speech spectrum under a fixed band-importance procedure.

Speech Perception in Noise (SPIN).

A sentence test contrasting predictable and unpredictable final words to isolate a listener's use of context.

Speech reception threshold (SRT).

The signal level or signal-to-noise ratio at which a listener recognizes 50% of the material; estimated efficiently by an adaptive staircase.

Speech Transmission Index (STI).

A 0-to-1 predictor of intelligibility derived from the modulation transfer function, standard for rating rooms and communication channels.

Key Researchers

Harvey Fletcher

(1884–1981). Acoustical physicist at Bell Telephone Laboratories whose research on hearing and speech perception produced the articulation-index framework that underlies modern intelligibility prediction. His program established that intelligibility could be computed from the audible fraction of the speech spectrum, the idea every later signal-based index refines. Wikipedia

Raymond D. Kent

(living). Speech scientist, emeritus at the University of Wisconsin–Madison, whose work on phonetic intelligibility testing in dysarthria established a systematic basis for measuring disordered-speech intelligibility and turning a global score into a diagnostic profile. ORCID - Google Scholar - Faculty page - Wikipedia

George A. Miller

(1920–2012). Cognitive psychologist whose 1951 study with Heise and Lichten showed that intelligibility rises sharply with the context and predictability of the test materials, a foundational demonstration of the top-down contribution to speech perception. Wikipedia - Wikidata

Jonathan E. Peelle

(living). Cognitive neuroscientist of speech perception and listening effort, at Northeastern University, whose synthesis of how the cognitive consequences of acoustic challenge are reflected in brain and behavior reframed intelligibility as one output of a resource-limited system. Google Scholar - Faculty page

M. Kathleen Pichora-Fuller

(living). Audiologist and cognitive-hearing scientist, Professor Emerita at the University of Toronto, who led the Framework for Understanding Effortful Listening, linking hearing impairment to cognitive load and recasting listening as effortful resource deployment. Google Scholar - Faculty page

Frequently Asked Questions

What is the difference between speech intelligibility and audibility?

Audibility is whether a sound can be heard at all, whether it exceeds the listener's threshold and any masking noise. Intelligibility is whether its linguistic content can be recognized. Speech can be fully audible yet unintelligible when noise, distortion, or reverberation has destroyed the cues that distinguish phonemes, which is why loudness alone does not guarantee understanding (French & Steinberg, 1947).

How is speech intelligibility measured?

By presenting calibrated material (nonsense syllables, phonetically balanced words, or sentences) and scoring the proportion recognized correctly. The full result is a psychometric function of signal-to-noise ratio, usually summarized by the speech reception threshold, the point of 50% correct, estimated with an adaptive staircase (Plomp & Mimpen, 1979).

What is the articulation index?

A predictive measure, between 0 and 1, that forecasts intelligibility from the signal alone by weighting the audible, undistorted fraction of the speech spectrum across frequency bands and summing (French & Steinberg, 1947). Kryter standardized its calculation (Kryter, 1962), and its modern successor is the Speech Intelligibility Index.

What is the Speech Transmission Index?

A single 0-to-1 predictor of intelligibility in a room or over a channel, derived from the modulation transfer function, a measure of how faithfully the channel preserves the slow amplitude fluctuations that carry speech (Houtgast & Steeneken, 1985). Unlike the articulation index, it captures reverberation and nonlinear distortion, so it is the standard for auditoria and public-address systems.

Why does context make speech easier to understand?

Because listeners use linguistic knowledge to predict upcoming words, so a predictable word can be recognized from partial acoustic information that would be insufficient in isolation. Miller, Heise, and Lichten showed that intelligibility rises sharply with context, which is why sentence tests score far higher than isolated-word tests at the same signal-to-noise ratio (Miller, Heise, & Lichten, 1951).

What is the speech reception threshold?

The signal level or signal-to-noise ratio at which a listener recognizes 50% of the speech material. It condenses a whole psychometric function into one interpretable number and is estimated efficiently by an adaptive procedure that tracks the threshold across trials (Plomp & Mimpen, 1979). It is a core clinical audiology measure.

Can a person understand speech perfectly and still find it effortful?

Yes. Recognition accuracy and the cognitive effort it costs are separable: a listener can score 100% while deploying substantial attention and working memory, a cost measurable through pupillometry, dual-task performance, and self-report even when the intelligibility score is at ceiling (Peelle, 2018; Pichora-Fuller et al., 2016).

How is intelligibility used in the speech clinic?

As an outcome measure in motor-speech disorders, where it quantifies how a disorder degrades communication and how therapy restores it. Phonetic scoring identifies which sound contrasts a disordered talker fails to convey, turning a global percentage into a diagnostic profile that points at the underlying impairment (Kent, Weismer, Kent, & Rosenbek, 1989).

References

French, N. R., & Steinberg, J. C. (1947). Factors governing the intelligibility of speech sounds. The Journal of the Acoustical Society of America, 19(1), 90–119. https://doi.org/10.1121/1.1916407

Houtgast, T., & Steeneken, H. J. M. (1985). A review of the MTF concept in room acoustics and its use for estimating speech intelligibility in auditoria. The Journal of the Acoustical Society of America, 77(3), 1069–1077. https://doi.org/10.1121/1.392224

Kalikow, D. N., Stevens, K. N., & Elliott, L. L. (1977). Development of a test of speech intelligibility in noise using sentence materials with controlled word predictability. The Journal of the Acoustical Society of America, 61(5), 1337–1351. https://doi.org/10.1121/1.381436

Kent, R. D., Weismer, G., Kent, J. F., & Rosenbek, J. C. (1989). Toward phonetic intelligibility testing in dysarthria. Journal of Speech and Hearing Disorders, 54(4), 482–499. https://doi.org/10.1044/jshd.5404.482

Kryter, K. D. (1962). Methods for the calculation and use of the articulation index. The Journal of the Acoustical Society of America, 34(11), 1689–1697. https://doi.org/10.1121/1.1909094

Miller, G. A., Heise, G. A., & Lichten, W. (1951). The intelligibility of speech as a function of the context of the test materials. Journal of Experimental Psychology, 41(5), 329–335. https://doi.org/10.1037/h0062491

Peelle, J. E. (2018). Listening effort: How the cognitive consequences of acoustic challenge are reflected in brain and behavior. Ear and Hearing, 39(2), 204–214. https://doi.org/10.1097/AUD.0000000000000494

Pichora-Fuller, M. K., Kramer, S. E., Eckert, M. A., Edwards, B., Hornsby, B. W. Y., Humes, L. E., Lemke, U., Lunner, T., Matthen, M., Mackersie, C. L., Naylor, G., Phillips, N. A., Richter, M., Rudner, M., Sommers, M. S., Tremblay, K. L., & Wingfield, A. (2016). Hearing impairment and cognitive energy: The Framework for Understanding Effortful Listening (FUEL). Ear and Hearing, 37(Suppl 1), 5S–27S. https://doi.org/10.1097/AUD.0000000000000312

Plomp, R., & Mimpen, A. M. (1979). Improving the reliability of testing the speech reception threshold for sentences. Audiology, 18(1), 43–52. https://doi.org/10.3109/00206097909072618