Can word duration alone cue which Mandarin homophone was heard?

Why duration might carry word identity

Speech carries timing information at several levels. Segment duration can distinguish words in languages with phonological length, and vowel duration helps cue contrasts in languages without it. Whether duration also marks the identity of a whole word, rather than only the sounds inside it, is a separate question. The paper treats word duration as a test case for that question.

The production literature gives a reason to expect such a link. Across languages, including Mandarin Chinese, high-frequency words tend to be shorter than low-frequency words. If listeners have stored that association, a shorter token could point toward the more frequent member of a homophone pair. The paper's experiments ask whether listeners actually use this mapping when duration is the only difference available.

Homophones isolate duration from every other cue

Mandarin was chosen because it has many disyllabic homophones: word pairs that share the same segments and the same tone but differ in meaning and frequency. Because the two members of a pair are segmentally and tonally identical, any systematic acoustic difference between them cannot be attributed to their sounds. That makes homophones a clean test bed for a word-level cue.

The design went further by removing other supports for recognition. Target words were embedded in a neutral carrier phrase and masked with noise, so duration was the only difference between the two members of a pair. Participants heard the pair and chose which order they had heard, using two buttons that displayed the homophone pair in opposite frequency orders. If duration carries word identity, listeners should map the shorter stimulus to the higher-frequency word.

Three accounts make different predictions

The paper sets up three explanations and uses them to interpret the results. A context integration account holds that duration matters indirectly, for example by giving listeners more time to bring sentential context to bear. A spectral integration account holds that duration shapes how listeners interpret the surrounding acoustic signal. A direct cueing account holds that word duration is itself associated with word identity in the lexicon.

The predictions diverge sharply in this design. Context integration and spectral integration both predict chance performance when context and spectral cues are held constant or removed. Only direct cueing predicts above-chance behavior. That contrast is what makes the homophone task informative rather than merely suggestive.

Shorter first word shifted choices toward high frequency

Experiment 1 tested 16 target homophone pairs selected from the Chinese Lexical Database. Each pair had exactly two members, both disyllabic, matched in part of speech, with the dominant part of speech covering at least 90% of occurrences and both members familiar to more than 80% of participants in a lexical decision task. High-frequency words had Zipf values from 3.8 to 4.8 and low-frequency words from 2.0 to 3.0, a mean difference of 1.58.

Stimuli were recorded by one male native Mandarin speaker. The high-frequency member of each pair was manipulated to be 25% shorter than its low-frequency counterpart, approximating the expected frequency effect on duration. In the critical analysis, auditory stimulus type significantly predicted the chosen order (β = 0.24, SE = 0.08, z = 3.01, p = .003): when the first word was shorter than the second, participants were more likely to select the high-low frequency ordering. Trial order did not predict responses (p = .793).

Participant screening mattered for interpreting this result. Of 200 participants, 164 passed the LexTALE task using a cutoff of 0.73 raw accuracy and 28 corrected accuracy, and the reported analysis concerns those remaining participants. Filler trials confirmed that the noise manipulation affected perception: accuracy on non-homophone filler pairs was 96.7% for Prolific and 95.3% for Xiaohongshu participants at low noise, falling to 82.7% and 75.2% at high noise.

The effect survived with no speech signal at all

Experiment 2 pushed the test further by removing the speech signal entirely. Target trials presented only noise at 65 dB, with no intelligible speech present, while preserving the duration difference between the high-frequency and low-frequency members of each pair. Listeners could therefore rely on nothing but timing.

The effect held. Participants still mapped the shorter duration to the higher-frequency word, and the paper reports that the magnitude of the effect did not differ reliably between the two noise conditions. This is the strongest evidence in the paper for direct cueing, because neither context integration nor spectral integration predicts above-chance performance when there is no signal to integrate.

Duration as stored lexical detail

The authors interpret the pattern as evidence that word duration can directly cue word identity and may be a phonetic detail incorporated into lexical representation. On this reading, listeners' mental lexicon encodes word-level temporal information, and that information is available during lexical access. The paper connects this to hybrid models of speech perception that combine abstract phonological categories with fine phonetic detail, with word duration as one dimension of that detail.

This is the authors' interpretation of a behavioral choice pattern, not a direct observation of stored representations. The experiments show that listeners can use duration when it is the only distinguishing cue. They do not show how duration combines with context and spectral information in natural speech, and the paper explicitly notes that the three accounts are not mutually exclusive in ordinary listening.

A modest effect from a narrow stimulus set

The reported effect was modest, and the stimuli came from a single speaker, 16 item pairs, and one fixed duration ratio of 25%. Those constraints bound how far the finding generalizes. The paper calls for testing graded duration differences, multiple speakers, and other languages before the mechanism's generality can be established.

The task also differs from ordinary listening in a specific way: participants chose between two printed orders, and the paper notes that simultaneous presentation of response options and audio can increase a positional bias favoring the first compatible orthographic form. The design controls for context and spectral cues precisely to isolate direct cueing, which means it does not measure how much duration contributes when those other cues are present.

Why the no-signal condition is the key comparison

The paper's most decision-relevant comparison is between Experiment 1 and Experiment 2. Partial masking leaves some signal available, so a positive result there could still be reconciled with indirect accounts. Removing the speech signal while preserving the duration difference eliminates that escape route, and the effect persisted. That is why the no-signal condition, not the masked condition alone, carries the argument for direct cueing.

For readers weighing whether word-level timing belongs in models of spoken word recognition, the paper's contribution is this specific demonstration: a frequency-linked duration difference, applied to segmentally identical words, shifted listeners' choices in the predicted direction even when nothing else was audible. The open question the paper leaves is not whether duration can cue word identity in this setup, but how strongly that cue operates alongside context and spectral information in natural speech.

Sources used in this answer