Co-occurrence of reduced word forms in natural speech

Malte C. Viebahn, Mirjam Ernestus, James M. McQueenInterspeech 2012

TL;DR

This study investigates whether reduced word forms in Dutch tend to appear together in natural speech, even after accounting for speech rate. The findings suggest that reduced forms do co-occur, offering insights into how our brains store and process language, potentially favoring episodic memory models over purely abstract ones.

What Problem Does This Paper Address?

When we speak spontaneously, our words often sound different from how they're written or how we might pronounce them in careful speech. This phenomenon, known as speech reduction, can involve shortening or even omitting sounds. For instance, the Dutch word "geweest" (been) might be pronounced as "g'weest." A key question in understanding natural speech production is whether these reductions are random or if they cluster together. Specifically, do speakers who reduce one word tend to reduce the next word as well? This paper tackles this by examining the co-occurrence of reduced word forms in Dutch, exploring how factors like time lag between words and their phonetic similarity might influence this tendency. Understanding this co-occurrence can help us differentiate between theories of how words are stored in our mental lexicon: abstractionist theories suggest we store abstract representations, while episodic theories propose we store detailed memories of each word's occurrence.

Methods

The researchers employed a corpus-based approach, analyzing spontaneous Dutch speech. They focused on past participles, a common word type that frequently exhibits reductions. Two types of reduction were specifically examined: schwa reduction in prefixes and /t/ reduction in suffixes.

The core of their method involved statistical modeling using linear mixed-effects models. These models allowed them to investigate the influence of several predictor variables on the presence and duration of reduced sounds. Key predictor variables included:

  • Previous Segment Realization: Whether the same sound (schwa or /t/) was reduced in the preceding word.
  • Time Lag: The interval between the offset of the first word and the onset of the second.
  • Phonological Relatedness: Whether the two words were identical, different but shared the same prefix, or were entirely different.

To control for confounding factors, they also included register (spontaneous, interview, read speech) and speech rate as control variables. For duration measurements, they used the natural logarithm to achieve more normal distributions.

Datasets

The study utilized data from two Dutch speech corpora:

  • Ernestus Corpus of Spontaneous Dutch (ECSD): Provided spontaneous speech data.
  • Corpus of Spoken Dutch (Corpus Gesproken Nederlands, CGN): Contributed data from both interview and read speech registers.

A total of 3,241 Dutch past participles were extracted from these corpora. The authors note that their analyses were based on the output of an automatic speech recognition system, and they cite previous work showing that the agreement between this system and human transcribers is comparable to inter-transcriber agreement.

Key Insights and Ideas

  • Reduced Forms Tend to Co-occur: The most striking finding is that reduced word forms are more likely to appear in sequence, even when controlling for speech rate. This suggests a systematic tendency in natural speech production.
  • Schwa Reduction is Influenced by Proximity and Identity: For schwa duration, the study found a significant interaction between the duration of the previous schwa, time lag, and relatedness. Specifically, schwa durations of succeeding tokens were similar when the tokens were identical words and occurred closely together in time.
  • /t/ Reduction Shows a More Complex Pattern: The findings for /t/ reduction were more varied. While there was a negative correlation between /t/ durations of succeeding past participles, the interaction with time lag and relatedness was less clear-cut, potentially due to dataset size.
  • Episodic vs. Abstractionist Models: The co-occurrence of reduced forms, particularly the influence of previous word occurrences, lends support to episodic models of lexical representation, which posit that fine-grained phonetic details of word occurrences are stored. Abstractionist models, which suggest abstract representations, would need to account for these findings through mechanisms like priming.

Key Results

  • Schwa Presence: Schwas were less likely to be present at higher speech rates and less likely in spontaneous speech compared to read or interview speech.
  • Schwa Duration: A significant three-way interaction was found: the duration of the previous schwa, time lag, and relatedness all influenced the current schwa duration. For identical words occurring closely together, succeeding schwa durations were similar.
  • /t/ Presence: Fewer /t/s were found at higher speech rates and in spontaneous speech compared to read speech.
  • /t/ Duration: Shorter /t/s were associated with longer /t/s in the previous participle. /t/s were also shorter when the preceding word was identical compared to when it was different. An interaction between time lag and relatedness was observed, with different words showing shorter /t/s at longer lags, and identical words showing longer /t/s at longer lags.

Limitations and Future Work

  • Dataset Size for /t/ Reduction: The authors acknowledge that the dataset for /t/ reduction was substantially smaller than for schwa reduction. This limited power might explain the absence of certain significant interactions, particularly regarding phonological relatedness.
  • Automatic Speech Recognition: While the authors note the high agreement of their ASR system with human transcribers, it's still an automated process, which can introduce its own set of potential errors or biases.
  • Future Directions: The paper suggests that the observed co-occurrence patterns could be explained by either episodic memory models (storing multiple phonetic variants) or abstractionist models (relying on priming of post-lexical processes). Further research could explore these theoretical implications more deeply, perhaps by manipulating experimental conditions to isolate the effects of priming versus stored episodic traces. Investigating other types of reductions or in different languages could also provide broader insights.

Summary updated

BibTeX
@inproceedings{viebahn2012co,
  author = {Malte C. Viebahn and Mirjam Ernestus and James M. McQueen},
  title = {{Co-occurrence of reduced word forms in natural speech}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure