Word-final [t]-deletion: an analysis on the segmental and sub-segmental level

Barbara Schuppler, Wim van Dommelen, Jacques Koreman, Mirjam ErnestusInterspeech 2009

TL;DR

This paper investigates the reduction of word-final /t/ sounds in conversational Dutch, revealing that both the frequency of word combinations and the surrounding sounds significantly influence whether the /t/ is pronounced. By analyzing both the presence/absence of the /t/ and its detailed acoustic properties, the study highlights the importance of sub-segmental information for understanding speech reduction and improving automatic speech recognition systems.

What Problem Does This Paper Address?

Conversational speech is often "reduced" compared to its canonical pronunciation, with sounds being deleted or altered. This phenomenon is crucial for understanding speech perception and production, and for building more accurate automatic speech recognition (ASR) systems. While previous research has explored factors influencing sound reduction, quantitative studies on standard Dutch word-final /t/-deletion have been limited, often focusing only on the segmental level (whether the /t/ is present or absent). This paper aims to fill this gap by providing a detailed analysis of word-final /t/-deletion in standard Dutch, examining both segmental and sub-segmental (acoustic) levels, and identifying the linguistic and acoustic factors that predict this reduction.

Methods

The study employs a two-pronged approach:

  • Study 1: Segmental Level Analysis: This study used a large corpus of conversational Dutch and an Automatic Speech Recognition (ASR) system to generate phonemic transcriptions. Statistical modeling, specifically mixed-effects logistic regression, was used to identify predictors for the presence or absence of word-final /t/s. Factors analyzed included bigram frequency (frequency of word pairs), surrounding phonetic segments (vowels, consonants, silence), word class (function vs. content words), and morphological class.
  • Study 2: Sub-segmental Level Analysis: This study focused on a smaller subset of the data, with manual annotations of acoustic properties of word-final /t/s. Researchers meticulously analyzed features like closure (present/absent, filled/silent), burst (present/absent, number, strength), and alveolar friction (presence, voicing, onset). Mixed-effects logistic regression was again used to model the acoustic variations based on linguistic factors (previous/following segments, word class, frequency) and acoustic properties themselves.

Datasets

  • CORPUS ERNESTUS: The primary dataset consists of six spontaneous Dutch dialogues from the CORPUS ERNESTUS. This corpus comprises approximately 90 minutes of speech per conversation, totaling 89,716 word tokens and 6,296 word types, spoken by 12 male native speakers of standard Dutch with academic degrees.
  • CGN (Corpus Gesproken Nederlands): Lexical frequency and bigram frequency data were extracted from the CGN.
  • Study 2 Subset: A subset of 130 word tokens (65 word types) from Study 1 was used for detailed sub-segmental acoustic analysis.

Key Insights and Ideas

  • Bigram Frequency is King (for Segmental Deletion): The frequency of a word pair (bigram) is a stronger predictor of word-final /t/ deletion than individual word frequency. This suggests that speakers anticipate upcoming words and adjust their articulation accordingly.
  • Context Matters, A Lot: Both the preceding and following sounds significantly influence /t/ deletion. /t/s are more likely to be deleted when surrounded by consonants compared to vowels or silence.
  • Sub-segmental Variation is Rich: Even when a /t/ is considered "absent" segmentally, there's a wide range of acoustic realizations. "Canonical" /t/s (with clear closure, burst, and alveolar friction) are surprisingly rare (11.5%), and complete absence is also infrequent (5.4%).
  • Frequency Drives Acoustic Reduction Too: Higher lexical frequency predicts less prominent bursts and more "filled" closures (with friction or nasality), indicating that frequent words undergo more acoustic reduction at the sub-segmental level.
  • ASR Needs More Than Just Phonemes: The significant difference between perceptually classified absent /t/s (25.4%) and automatically transcribed absent /t/s (40.8%) suggests that current ASR systems, often trained on read speech, miss crucial nuances of spontaneous speech reduction. Incorporating sub-segmental information could improve ASR accuracy.

Key Results

  • Segmental Deletion: 35.15% of word-final /t/s were classified as absent.
  • Predictors for Absence: Bigram frequency (negative correlation), preceding consonants (higher absence), following consonants (higher absence), and function word status (though less impactful than bigram frequency) were significant predictors of /t/ deletion.
  • Acoustic Properties:
    • Higher lexical frequency predicted less probable bursts.
  • Filled closures were more frequent in higher-frequency words.
  • Alveolar friction was more likely when /t/ followed a vowel.
  • Multiple bursts were more likely before homorganic consonants.
  • ASR vs. Auditory Classification: Agreement was 75.4% on the presence/absence of /t/s between automatic transcription and human auditory classification, with automatic systems identifying more absent /t/s.

Limitations and Future Work

  • ASR Training Data: The study notes that ASR systems trained on read speech may not accurately reflect spontaneous speech. Future work could involve training ASR models on spontaneous speech data to improve their handling of reduced forms.
  • Sub-segmental Analysis Scope: While Study 2 provided detailed acoustic analysis, it focused on a limited segmental context (preceded by vowel/nasal, followed by fricative). Expanding this analysis to a wider range of contexts would be beneficial.
  • Improving ASR: The authors suggest that ASR systems could be improved by first identifying potential /t/ realization intervals and then applying detailed analysis and classification techniques informed by sub-segmental acoustic data. This could lead to more accurate automatic transcriptions of spontaneous speech.

Summary updated

BibTeX
@inproceedings{schuppler2009word,
  author = {Barbara Schuppler and Wim van Dommelen and Jacques Koreman and Mirjam Ernestus},
  title = {{Word-final [t]-deletion: an analysis on the segmental and sub-segmental level}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure