PaperlayerView exact revision on arXiv

Source-linked AI summary

Direct cueing of word identity through word duration: evidence from Mandarin homophones

Xiaomeng Zhu, Megan Niu, Jason Shaw, Kevin Tang

arXiv:2610.30682v1cs.CL

TL;DR

Whether word duration directly cues word identity, rather than only affecting access to context or other acoustic cues, remains an open question. The paper tests Mandarin homophones with duration isolated from other cues and finds that listeners associate shorter stimuli with higher-frequency words, including when no speech signal is available.

  • Problem

    It remains unclear whether word duration directly cues word identity or only affects recognition indirectly through context or processing of other cues.

    Source: §I · pp. 3–4

  • Method

    The authors tested homophone-order choices in two experiments, manipulating duration while controlling other cues and then testing the effect with noise alone.

    Source: A. Methods · p. 18, §V · p. 24

  • Results

    Listeners more often placed the higher-frequency homophone first when its auditory stimulus was shorter, including when only duration-matched noise was available.

    Source: B. Results · pp. 19–20, §V · p. 24

  • Takeaways & Limitations

    The findings are most consistent with word duration directly cueing identity and suggest that lexical representations encode word-level temporal detail available for lexical access.

    Source: Because the design removed sentential… · p. 24

Abstract

from arXiv · show

While there is increasing evidence that words are stored in the mental lexicon with some phonetic detail, there are open questions about exactly which phonetic details are stored and whether they are functionally relevant for spoken word recognition. Word-level phonetic cues, such as word duration, provide an appropriate test case. Across languages, including Mandarin Chinese, high frequency words are shorter than low frequency words. Using a homophone ordering task, we investigate whether differences in word duration impact word recognition. Mandarin Chinese was examined since it has abundant disyllabic homophones. Our results indicate that participants were able to use word duration to recognize the order of homophone pairs, selected to vary in lexical frequency. Both with high noise levels and when no other cues were available, participants mapped the shorter duration auditory stimulus to the higher frequency word. The results provide evidence that word duration can directly cue the identity of words, suggesting that it may be a phonetic detail incorporated into the lexical representation. We discuss this direct cueing account alongside alternative accounts of how word duration may modulate word recognition.

I. INTRODUCTION

Duration is known to cue phonological units, but whether word duration directly helps listeners identify words remains unresolved. The paper tests this with Mandarin homophones, where duration can be isolated from segmental identity and compared with indirect accounts of recognition.

I. INTRODUCTION

  • Duration cues phonological contrasts in languages with and without phonological length, while its role in identifying higher-level units such as words is less clear.The introduction contrasts segment-level duration cues with the open question about word-level cueing.Source: §I · p. 3
  • Prior findings leave open whether duration directly cues word identity or affects recognition indirectly through context or processing of other cues.Reported perceptual effects vary by task and context, and prior evidence does not distinguish direct cueing from indirect accounts.Source: §I · pp. 3–4

Whether listeners use such differences to recognize words and to distinguish homophones

tion. This suggests that listeners can exploit acoustic details of homophones. Furthermore,

  • The study predicts that listeners will map shorter homophone tokens to higher-frequency words and longer tokens to lower-frequency words.It contrasts direct cueing with context integration and spectral integration, which predict no reliable association when context and other acoustic cues are unavailable.Source: tion. This suggests that listeners can… · pp. 6–8

II. EXPERIMENT 1: SOME SIGNAL FOR HOMOPHONE PAIRS

Experiment 1 tests whether listeners use duration alone to order frequency-differing Mandarin homophones. The materials and task isolate duration while selecting familiar, disyllabic pairs that differ in lexical frequency.

1. Design

  • Participants hear homophone pairs whose only target-signal difference is duration, then choose which of two frequency orders they heard.The response options are high-low and low-high; the expected mapping is shorter-to-higher-frequency and longer-to-lower-frequency.Source: §1 · p. 9

2. Materials

  • The task contained 48 trials: 16 homophone targets and 32 fillers, with phonetically similar non-homophone filler pairs.The interface asked participants which word order they heard and displayed the two frequency orders as response options.Source: §2 · pp. 9–11, Figure 1 · p. 10
  • The 16 target pairs were selected from eligible two-member disyllabic homophone sets, with matching part of speech and familiarity criteria.The selection also excluded near-synonyms by requiring cosine similarity below 0.5.Source: §2 · pp. 10–11
  • Target pairs differed in lexical frequency: high-frequency words had Zipf values of 3.8–4.8, low-frequency words 2.0–3.0, with a mean difference of 1.58.The target set was selected from 143 eligible homophone pairs.Source: §2 · p. 11
  • The high-frequency token was manipulated to be 25% shorter than its low-frequency counterpart, while other recorded signal properties were preserved.Tokens were chosen from six recorded repetitions after excluding unusually long low-frequency productions.Source: §2 · p. 12

and Weenink, 2021) Manipulation Object through the parselmouth library (Jadoul et al.,

The experiment masked target words with duration-matched noise and then presented participants with a single-play ordering task. Participants also completed filler trials and lexical and demographic measures.

and Weenink, 2021) Manipulation Object through the parselmouth library (Jadoul et al.,

  • Speech-shaped noise matched each manipulated word’s duration, and Experiment 1 targets used the low-noise 0 dB condition.The materials also included a medium-noise −8 dB condition, intended to make word identification more difficult.Source: and Weenink, 2021) Manipulation Object… · pp. 12–13

3. Procedure

  • Participants previewed both response options before hearing the stimulus, which was intended to encourage attention to both alternatives.The options remained visible during playback, and their positions were randomized across trials.Source: §3 · p. 13
  • Participants heard each audio stimulus once, with no replay option, and advanced after playback and the question appeared.The inter-trial interval was 500 ms.Source: §3 · pp. 13–14
  • The 32 filler trials varied across eight conditions and were assigned either low or medium noise; practice included one trial at each noise level.The easier filler trials were intended to encourage participants through the task.Source: §3 · p. 14
  • Participants completed a Mandarin LexTALE task and a demographic questionnaire after the homophone-identification task.Participants were recruited through Prolific and Xiaohongshu.Source: §3 · pp. 13–14

In the last part of the experiment, participants completed a demographics questionnaire

After screening participants, the authors analyzed homophone-order choices with mixed-effects logistic regression. Duration pattern predicted choices in Experiment 1, and the next experiment tested the effect when no speech signal was available.

The dependent variable was participants’ choice of homophone order, coded as either

  • The analysis modeled binary homophone-order choices with stimulus type and standardized trial number as fixed effects and participant and item random intercepts.Low-high was the reference response level, and only target trials from participants who passed lexical screening were included.Source: The dependent variable was… · pp. 15–16

We first analyzed the performance of participants on the supplementary LexTALE task

  • 164 of 200 participants passed the LexTALE proficiency screen, using cutoffs of 0.73 raw accuracy and 28 corrected accuracy.The screening was intended to exclude participants with insufficient Mandarin proficiency for the experiment.Source: We first analyzed the performance of… · p. 16
  • Filler accuracy was higher in low noise than high noise, and both Prolific and Xiaohongshu groups showed comparable filler performance.Low-noise accuracy was 96.7% for Prolific and 95.3% for Xiaohongshu; high-noise accuracy was 82.7% and 75.2%, respectively.Source: We first analyzed the performance of… · p. 16
  • The authors fitted the mixed-effects logistic model to test whether stimulus duration pattern predicted homophone-order choice.The model included standardized trial number as a fixed predictor and random intercepts for participant and target item; results were reported in Table II.Source: We first analyzed the performance of… · p. 17, Table II · p. 17
  • Auditory stimulus type significantly predicted homophone-order choice (β = 0.24, SE = 0.08, z = 3.01, p = .003).Participants more often selected high-low when the first homophone was shorter than the second; trial order was not significant (p = .793).Source: We first analyzed the performance of… · p. 17

III. EXPERIMENT 2: NO SIGNAL FOR HOMOPHONE PAIRS

  • Experiment 1 showed that duration influenced homophone-order choices even though the paired words were segmentally identical and differed only in manipulated duration.Participants were more likely to select high-low when stimuli were short-long.Source: §III · p. 18
  • Experiment 2 tested whether the duration effect persisted without an intelligible speech signal, using the highest-noise condition.The noise-only target condition had an intensity of 65 dB while preserving the duration difference between homophone items.Source: §III · p. 18

A. Methods

Experiment 2 tested whether duration patterns predicted homophone-order choices under the highest-noise condition, and the results remained consistent with duration cueing word identity. The discussion argues that duration can provide a direct word-identity cue, while noting that context and spectral integration may also matter in naturalistic settings.

A. Methods

  • Experiment 2 used the highest noise condition, noise only at 65 dB, and recruited 200 additional participants.Of 200 participants, 174 passed LexTale; filler accuracy was 0.78 in high-noise and 0.95 in low-noise conditions.Source: A. Methods · p. 18, B. Results · p. 19

B. Results

  • Auditory stimulus type significantly predicted homophone-order choice in Experiment 2, with short-long stimuli making participants more likely to select high-low order.The effect did not reliably differ across noise conditions, and the authors report detecting it even without an intelligible signal.Source: A. Methods · p. 18, B. Results · pp. 19–20
  • The auditory-stimulus-type by noise-level interaction was not significant (p = .700), so the duration-pattern effect did not reliably differ across noise conditions.The authors interpret the significant stimulus-type effect alongside this interaction as evidence that duration was detectable even without an intelligible signal.Source: B. Results · pp. 19–20, Table IV · p. 20

IV. DISCUSSION

  • The discussion distinguishes context integration, spectral integration, and direct cueing as three ways duration could modulate word recognition.Context integration allows more time for sentential prediction, spectral integration allows more time to process other identity cues, and direct cueing treats duration itself as an identity cue.Source: Generally speaking, increasing the… · p. 21

Generally speaking, increasing the duration of an interval could facilitate word recognition

  • Listeners mapped the shorter stimulus to the higher-frequency homophone, and both experiments showed above-chance ordering when context and other phonetic cues were unavailable.The authors argue that this pattern supports direct cueing beyond merely allowing more time for context or spectral processing.Source: Generally speaking, increasing the… · pp. 21–23
  • The discussion does not rule out context and spectral integration; in naturalistic contexts, longer duration may also provide time to process contextual and spectral cues.The paper’s experiments used a neutral context and withheld other cues, so their results do not settle how these accounts operate in naturalistic contexts.Source: Generally speaking, increasing the… · pp. 21–22
  • The findings suggest that word duration, a phonetic detail not represented in strings of segments, may be incorporated into lexical representation.The authors note that the precise mechanism and dimensions of phonetic detail encoded in lexical representations remain open issues.Source: Generally speaking, increasing the… · p. 23

V. CONCLUSION

In Mandarin homophone ordering, listeners mapped the shorter item to the higher-frequency word under masking and even when only duration-matched noise was presented. The pattern favors direct word-duration cueing and supports word-level timing in lexical representations, though its generality remains to be established.

V. CONCLUSION

  • Listeners more often placed the higher-frequency homophone first when its stimulus was shorter, both under partial masking and with duration-matched noise but no speech signal.The effect magnitude did not differ reliably between the noise conditions.Source: §V · p. 24

Because the design removed sentential context and held all non-durational cues constant,

  • Because context integration and spectral integration alone poorly explain the results, the authors favor direct cueing in which duration is associated with word identity.They infer that word-level temporal detail is encoded in lexical representations and available for lexical access.Source: Because the design removed sentential… · p. 24
  • The findings extend phonetic-detail accounts beyond segment-level cues: spoken-word models need to represent whole-word timing alongside segmental and tonal form.The authors raise pitch and intensity as possible additional word-specific properties available to listeners.Source: Because the design removed sentential… · p. 24
  • Generality remains unestablished because the effects were modest and the stimuli used one speaker, 16 item pairs, and one fixed duration ratio.The authors call for tests with graded duration differences, multiple speakers, and other languages, as well as in natural speech with context and spectral cues.Source: Because the design removed sentential… · p. 25

APPENDIX A: STIMULUS SELECTION

The passage reports selecting 16 target pairs using Zipf-scale word frequencies and describes an exhaustive search over measures for each pair.

APPENDIX A: STIMULUS SELECTION

  • The 16 target pairs were selected using Zipf-scale word frequencies, which ranged from 1 to 7, through an exhaustive search over measures computed for each pair.Source: Appendix A · p. 26

3. Maximization of the mean frequency difference across all items (the mean frequency

The stated measure subtracts the mean frequency of low-frequency words from the mean frequency of high-frequency words.

3. Maximization of the mean frequency difference across all items (the mean frequency

  • The stated measure subtracts the mean frequency of all low-frequency words from the mean frequency of all high-frequency words.Source: §3 · p. 27

4. Maximization of the minimum frequency difference (the minimum frequency of all

The stated measure subtracts the maximum frequency of low-frequency words from the minimum frequency of high-frequency words.

4. Maximization of the minimum frequency difference (the minimum frequency of all

  • The stated difference subtracts the maximum frequency of all low-frequency words from the minimum frequency of all high-frequency words.Source: §4 · p. 27

Norris, D. (1994). “Shortlist: A connectionist model of continuous speech recognition,”

This reference block lists work on spoken-word recognition, lexical ambiguity, acoustic detail, and phonetic information in the lexicon.

Norris, D. (1994). “Shortlist: A connectionist model of continuous speech recognition,”

  • The listed recognition models include Norris and McQueen’s “Shortlist B” and Scharenborg’s work on durational information in spoken-word recognition.The passages provide titles and bibliographic details, not findings from these works.Source: Norris, D. (1994). “Shortlist: A… · pp. 33–34
  • Other listed works concern word-specific phonetics, articulatory planning, and the importance of acoustic properties in spontaneous-speech perception.The passages identify these topics in the cited titles.Source: Norris, D. (1994). “Shortlist: A… · p. 33
  • Sanker’s cited studies address lexical ambiguity alongside acoustic details, acoustic distance, and prior exposure in homophone discrimination.These are title-level descriptions only.Source: Norris, D. (1994). “Shortlist: A… · pp. 33–34
  • The block also lists work on durational information in homophone disambiguation and acoustic features in speech recognition.The supplied citations identify these topics but do not report study outcomes.Source: Norris, D. (1994). “Shortlist: A… · pp. 33–34

tomatic disambiguation of homophones in spontaneous German,” Computer Speech

This reference block lists work on Chinese lexical resources, lexicalized intonation, contextual predictability, and prosody.

tomatic disambiguation of homophones in spontaneous German,” Computer Speech

(CLD): A large-scale lexical database for simplified Mandarin Chinese,” Behavior Research

This reference block covers phonetic detail, speech corpora, duration, prosody, and lexical effects in Mandarin and English.

(CLD): A large-scale lexical database for simplified Mandarin Chinese,” Behavior Research

Loading 2610.30682v1…