On the Use/Misuse of the Term ‘Phoneme’

Roger K. Moore, Lucy SkidmoreInterspeech 2019

TL;DR

This paper investigates how the term "phoneme" is used in the speech science and technology community, finding that a significant portion of researchers misuse it, often confusing it with "phone" or treating it as a concrete acoustic unit. This misuse can hinder understanding and limit the potential of speech models and technologies.

What Problem Does This Paper Address?

The term "phoneme" is fundamental to speech science and technology, representing the smallest unit of sound that distinguishes meaning in a language. However, the authors suspect that many researchers use it casually, not as a well-defined abstract concept but rather as a simple acoustic building block. This casual usage might stem from a misunderstanding of the crucial difference between "phonetic" (physical sound) and "phonemic" (abstract, meaning-distinguishing) levels of description, and the concept of "phonemic contrast." If this suspicion is true, it means a significant part of the community may be missing opportunities to fully understand and exploit the psychological phenomenon of phonemes, potentially limiting the explanatory power of their models and the capabilities of their technologies.

Methods

The authors conducted an empirical study by analyzing the accepted papers from the INTERSPEECH 2018 conference. They counted the occurrences of the term "phoneme" and examined the context in which it was used. To assess the accuracy of its usage, they looked for definitions provided within the papers. They also categorized papers into "speech science" and "speech technology" to see if usage differed between these domains. A historical comparison was made with papers from the ICSLP 1998 conference to observe trends over time.

Datasets

The primary dataset analyzed was the collection of 791 papers accepted for INTERSPEECH 2018. A secondary dataset was used for historical comparison: 831 papers from the 5th International Conference on Spoken Language Processing (ICSLP) in 1998. No new datasets were introduced or created for this study.

Key Insights and Ideas

  • Widespread Misuse: A significant portion of researchers (around 40%) in the INTERSPEECH 2018 proceedings used the term "phoneme" in a way that could be construed as misuse.
  • Confusion with "Phone": Many authors seemed unaware of the critical difference between "phonetic" (physical sound) and "phonemic" (abstract, meaning-distinguishing) levels of description.
  • "Beads-on-a-String" Fallacy: There's a tendency to treat phonemes as discrete, well-defined acoustic units with clear boundaries, which contradicts the reality of speech production where phonemic information can spread and boundaries are not always clear.
  • Technology vs. Science: While the term "phoneme" appears slightly more frequently in speech science papers, the rate of misuse was found to be similar across both speech science and speech technology domains, which was a surprising finding.
  • Lack of Definitions: Very few papers provided any definition of "phoneme," suggesting an assumption that its meaning is universally understood, which the study indicates is not the case.
  • Historical Consistency: The pattern of usage and the proportion of papers using the term "phoneme" have remained remarkably consistent over the past 20 years.

Key Results

  • In INTERSPEECH 2018, 34% of papers contained the term "phoneme," with an average of 7.69 occurrences per paper.
  • Around 40% of papers mentioning "phoneme" were found to misuse the term.
  • Of the papers with potentially inappropriate usage, 25% were categorized as "science" and 75% as "technology."
  • Only six papers out of 265 provided anything approaching a definition of "phoneme," and none were deemed satisfactory.
  • The term "phoneme" appeared in 66% of speech science papers and 65% of speech technology papers in INTERSPEECH 2018, indicating it was absent in a majority of papers in both categories.

Limitations and Future Work

The authors acknowledge that their analysis is based on the interpretation of usage within published papers and that some authors might have a correct understanding of the term but used it casually. They also note that the emergence of "end-to-end" systems raises questions about the role of explicit intermediate levels of representation.

The paper makes three key recommendations for the future:

  1. Researchers should avoid the term "phoneme" unless they are certain of its meaning, using "phone" for generic speech sounds and "phoneme" only for abstract, meaning-distinguishing units.
  2. Educators and supervisors should ensure newcomers are thoroughly briefed on the distinction between phonetic and phonemic levels, the significance of phonemic contrast, and correct usage.
  3. Community associations (like ISCA and IEEE) should take steps to raise awareness about the importance of using the term "phoneme" correctly.

Summary updated

BibTeX
@inproceedings{moore2019use,
  author = {Roger K. Moore and Lucy Skidmore},
  title = {{On the Use/Misuse of the Term ‘Phoneme’}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure