cs.CLAug 27, 2026

Letters hide the truth from our eyes: English homophones have meaningfully different phonetic realizations

Authors: Yu-Hsiang Tseng, Mirjam T. C. Ernestus, Louis F. M. ten Bosch, R. Harald Baayen

Organizations: University of Tübingen, Germany · Radboud University, the Netherlands

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The distribution of spoken word duration of English homophones is known to co-vary with frequency of use. This study investigates whether other aspects of the phonetic realization of homophones also differ. A series of quantitative investigations of 14,000 homophone tokens in American television news broadcasts revealed that the tokens of homophone pairs such as weight and wait have different phonetic realizations, and that these can be predicted from their meanings in utterance context. These systematic differences remain even when taking duration-related variation into account. Time-normalized spectrograms emerged as an excellent tool for probing the fine details of phonetic realization, and obviate the need for phonetic transcriptions, which inevitably hide the phonetic truth from our eyes.

Figures & tables

Figure 1: Spectrogram of a 470ms “wait” token (upper panel) and the corresponding fixed-size spectrogram with adaptive hop length (lower panel). The horizontal dashed lines indicate the frequency steps for the 21 Mel-frequency banks. The vertical dashed lines represent the 50 timesteps that are obtained with a step size of 9.4ms.
Figure 2: LDA classification accuracies for the speech tokens of the members of homophones within homophone pairs. The black vertical dashed line is the 50% baseline, the red vertical dashed line represents the overall mean accuracy with a 95% confidence interval in red, and the horizontal blue lines indicate 2 standard errors calculated with 10 cross-validation runs. Mean accuracies are represented by blue dots.
Figure 3: An example of a GAM for a homophone pair: wait vs. weight . (a) The averaged spectrogram of weight ; (b) the averaged spectrogram of wait ; (c) the phone probabilities of wait/weight in normalized time; (d) the difference spectrogram.
Figure 4: Contextualized embeddings of homophone tokens. Each point stands for one word token’s CE and is color-coded by word identity. All CEs are first transformed into word space using LDA and reduced into a two-dimensional space using t-SNE for visualization. Words are well-separated in the figure, consistent with the 95% word classification accuracy based on CEs.
Figure 5: Alignments of predicted meaning vectors (left panel) and form vectors (right panel) compared to a full permutation baseline (in red). The vertical dashed lines represent mean accuracies. Confidence intervals are based on 10-fold cross-validation.
Figure 6: RL2R_{\text{L2}} scores for comprehension (left panel) and production (right panel). Blue dots represent empirical RL2R_{\text{L2}} scores, red dots and red horizontal lines represent 95% credible intervals and means based on 30 within-word permutation cross-validation runs.
Figure 7: Selected difference spectrograms using GAMs applied to the observed spectra (contour lines) and using linear mappings from CE centroids (color coding). The close match between the predictions from LDL and the difference surfaces from GAM indicates that homophones’ prototypical meanings co-determine their prototypical spectra. Upper panels show phone alignment profiles as references. Four homophone twins are shown: (a) sell and cell , (b) son and sun , and (c) council and counsel .
Figure 8: R2R^{2} distributions across homophone pairs of regression models for duration. Predicting from embeddings (GPT2, blue) has a much lower R2R^{2} compared to predicting from the speech vectors derived from time-normalized spectrum (Spec, orange). Unsurprisingly, predicting from the fitted speech vectors (Predicted Spec, green) has the same overall R2R^{2} as predicting from embeddings. R2R^{2} is lowest when predicting from deflated speech vectors (Deflated Spec, red) and similar to the R2R^{2} obtained with permuted spectrograms (Perm Spec, brown). Deflating the speech vectors successfully removes meaningful duration effects.
Figure 9: Distributions (across the 35 homophone pairs) of LDA held-out accuracy (left panel) and nearest neighbor held-out accuracy (right panel) for models predicting from the observed spectrograms (orange), the deflated spectrograms (green), and from duration only (red). Removing duration from the spectrogram decreases prediction accuracy somewhat, but accuracy nevertheless remains high.

Explore similar work

CardsList

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

    Oct 22, 2026Xiaomeng Zhu, Megan Niu, Jason Shaw +1Speech Processing

    Department of Linguistics, Yale University, New Haven, USA · Department of English Language and Linguistics, Institute of English and American Studies, Faculty of Arts and Humanities, Heinrich Heine University Düsseldorf, Düsseldorf, 40225, Germany · Also at: Department of Linguistics, University of Florida, Gainesville, Florida, 32611-5454, United States of America

  2. Using embeddings to predict spoken word duration and pitch in Mandarin monosyllabic words

    Jul 2, 2026Xiaoyun Jin, Mirjam Ernestus, R. Harald BaayenSpeech ProcessingSpeech Prosody

    Quantitative Linguistik, University of Tuebingen, Germany · Center for Language Studies, Radboud University, 6525 HT Nijmegen, The Netherlands

  3. PHONOS: PHOnetic Neutralization for Online Streaming Applications

    Mar 27, 2026Waris Quamer, Mu-Ruei Tseng, Ghady Nasrallah +1Speech ProcessingSpeaker Anonymization

    Department of Computer Science & Engineering, Texas A&M University, College Station, US