Spoken word recognition

Spoken word recognition is the process of identifying familiar spoken forms from an unfolding speech signal and accessing the corresponding lexemes. It occurs as information arrives: the beginning of an utterance may fit several possible forms, while later sounds distinguish them. The temporary candidate set is a lexical cohort.

For example, the beginning of “candle” is also compatible with “candy”. As the later sounds arrive, the evidence changes. Recognizing the item therefore involves both sensitivity to the incoming sounds and access to stored phonological representations; it is not necessary to wait for a silent gap before beginning to interpret the signal.

A 2005 mouse tracking experiment made this changing evidence visible in action. Participants heard a target name while moving toward one of two pictures. A similar-sounding alternative altered the movement path even on trials ending at the correct picture. The result supported an account in which spoken word recognition can influence an ongoing action before the final choice is complete.^[ref]

The measured cursor path is a behavioural consequence of recognition, rather than a direct recording of a phonological representation. The inference linking the two must be made explicit.

A word's uniqueness point is the earliest point in its spoken form at which the sounds heard so far distinguish it from all competing words in a specified vocabulary. It connects a cohort model of spoken word recognition to the changing size of the lexical cohort.^[ref]

For an invented vocabulary, let three word forms be the phoneme sequences $pata$, $pana$ and $pita$, with spaces ignored and each symbol representing one phoneme. The third word becomes unique after $pi$, while the first two remain ambiguous after $pa$ and diverge at the following consonant. Uniqueness can therefore precede the end of a word.

This is a property of the chosen vocabulary and transcription, not a universal timestamp attached to a spelling. Adding another word beginning $pi$ changes the example. A word that is a complete prefix of another may require an end-of-word cue. Measured reaction time also includes perceptual, decision and motor processes, so it need not equal the lexical uniqueness time.

A mispronunciation detection task asks listeners to report an altered pronunciation in speech. The manipulation can vary where a change occurs within a word or how strongly the surrounding context predicts the word; responses can be scored for detection and reaction time.

In the 1978 spoken word recognition experiments, listeners in the detection condition saw the original passage while hearing it. The study compared this with a shadowing task, in which listeners repeated ongoing speech. Effects of syllable position appeared in detection, while contextual effects differed between the tasks.^[ref]

The written version supplies information unavailable to an ordinary listener without a transcript. Thus a quick correct response can reflect comparison with that text as well as processing of the incoming sounds. Keeping the altered sound, context and instructions explicit makes the task interpretable rather than treating detection accuracy as a pure measure of one isolated cognitive stage.

A shadowing task asks a participant to repeat incoming speech aloud while continuing to listen. The delay and accuracy of repetition provide behavioural measures of processing an unfolding speech stream. The task differs from repeating a whole sentence only after the speaker has finished.

William D. Marslen-Wilson and Alan Welsh used shadowing with altered pronunciations in continuous speech. Whether listeners repeated a changed form or fluently restored the expected word depended on aspects of the input and context. Comparison with a mispronunciation detection task helped show that task instructions change which influences appear in the measured response.^[ref]

A restoration is an observation about the spoken response. It does not alone establish that a listener consciously heard the original sound correctly, nor that every intervening processing stage had the same representation. This makes shadowing useful for testing the time course of spoken word recognition while keeping response production part of the explanation.

A cohort model explains spoken word recognition through an evolving set of candidate words, the lexical cohort. An initial stretch of speech activates plausible candidates; later acoustic evidence and, depending on the model version, contextual information constrain recognition. William D. Marslen-Wilson and Alan Welsh's 1978 study connected this account with experimental manipulations of continuous speech.^[ref]

For an illustrative English vocabulary, the opening sounds of “candle”, “candy” and “canister” initially overlap. Hearing the next consonant can remove “canister” while leaving the other two plausible; later sounds separate them. The candidates depend on pronunciation and the listener's vocabulary, not merely on matching written letters.

A deliberately simplified noise-free rule isHere $C_t$ is the candidate set after the evidence available at step $t$. This filtering illustration is not a fitted psychological model: noisy speech, graded activation and recovery from misleading evidence require additional assumptions. A uniqueness point concerns when the remaining acoustic prefix distinguishes a word within a specified vocabulary; it need not coincide with the listener's measured response time.

The original task comparisons motivate testing a model on intermediate behaviour and contextual effects, not just final accuracy.^[ref]

A lexical cohort is a set of candidate words compatible with an initial stretch of speech during spoken word recognition. In a cohort model, new acoustic information changes which candidates remain plausible. The term describes candidates in recognition, rather than a group of people followed in a longitudinal study.^[ref]

For an English example, an initial “cap” can fit “captain”, “captive” and “capsule”; later sounds distinguish them. The example assumes a particular vocabulary and pronunciation: spelling alone does not specify an acoustic cohort.

The 1978 experiments compared repeating ongoing speech with detecting altered pronunciations while following a written passage. Effects of where a mispronunciation occurred, and differences in the influence of context across tasks, constrained accounts of how incoming speech and prior expectations interact.^[ref] A candidate set is a model construct; behaviour supplies evidence about its time course rather than a literal inventory of consciously considered words.

Phonological competition occurs when speech forms with overlapping sounds are simultaneously plausible candidates during spoken word recognition and influence the eventual selection. For example, hearing the beginning of “candle” can make “candy” temporarily relevant as well. Later acoustic evidence can resolve the ambiguity even though the alternative has already affected processing.

In a 2005 experiment, a listener saw two pictures and used a mouse to select the named object. When the competing picture had a similar-sounding name, such as “candy” beside “candle”, the movement toward the correct object was drawn farther toward that competitor than toward an unrelated alternative. The comparison demonstrated an effect of phonological competition on the path of the response, beyond whether the final click was correct.^[ref]

This suggests a useful experimental distinction: accuracy measures the final choice, while reaction time and mouse tracking can reveal costs and intermediate influences. Similarity alone does not establish competition; a controlled manipulation must show that the alternative changes processing or behaviour.