Voicing profile of Polish sonorants: [r] in obstruent clusters

J. Sieczkowska, Bernd Möbius, Antje Schweitzer, Michael Walsh, Grzegorz DogilInterspeech 2009

TL;DR

This paper investigates the "voicing profile" of the Polish sonorant [r], analyzing how its voicing status changes frame-by-frame within different phonetic contexts. Using automated tools on a large speech corpus, they found that [r] devoices significantly, especially when word- or syllable-final and preceded by a voiceless stop, suggesting that both phonetic context and prosodic position play a role in its realization.

What Problem Does This Paper Address?

Understanding how speech sounds are produced and perceived is crucial for fields like speech synthesis, recognition, and language acquisition. While the concept of "voicing" (whether vocal cords vibrate) is a fundamental phonetic feature, its application to sonorants (like [r], [l], [m], [n]) is less straightforward than for obstruents (like [p], [t], [k]). Traditional methods often treat voicing as a binary (voiced/voiceless) or categorical feature. However, in continuous speech, the voicing status of a sound can vary dynamically. This paper addresses the need for a more nuanced understanding of sonorant voicing, specifically focusing on the Polish liquid [r], by analyzing its "voicing profile" – a frame-by-frame depiction of its voicing status – across various phonological and prosodic environments. This is particularly relevant for Polish, where sonorants can participate in voicing oppositions, unlike in many other languages.

Methods

The researchers employed an automated approach to extract voicing profiles. They utilized the ESPS get_F0 tool to obtain frame-by-frame voicing decisions at a sampling rate of 100 Hz. This yielded a voicing status for each frame, which could be interpolated if it fell between distinct voicing states. The study then analyzed these profiles for Polish sonorants ([j, l, w, m, n, ŋ, r]) within specific contexts, such as word-final position, syllable-final position, and preceding voiceless stops. They used the IMS German Festival synthesis system to extract features like consonant type, place of articulation, and voicing structure. The analysis focused on averaging voicing profiles across multiple exemplars within defined contexts to identify tendencies.

Datasets

The study utilized the BOSS corpora, specifically designed for Polish unit selection speech synthesis. This corpus comprises 115 minutes of speech from a professional speaker, including various phrase types designed to cover frequent consonant structures, diphones, triphones, and common vocabulary. The paper provides a table (Table 1) detailing the number of analyzed sonorants, with [r] having 3394 instances.

Key Insights and Ideas

  • Sonorants aren't always voiced: While sonorants are generally considered to be voiced due to their production mechanism, this study confirms that in Polish, the sonorant [r] can indeed devoice significantly.
  • [r] is the "devoicing star": Among the Polish sonorants analyzed, [r] exhibits the most variable voicing profile and is the primary sonorant to undergo devoicing in the contexts studied. Other sonorants were found to be categorically voiced.
  • Context matters, especially for [r]: The devoicing of [r] is strongly influenced by its phonetic environment. Specifically, it devoices considerably when it is word- or syllable-final and immediately preceded by a voiceless stop.
  • Prosody plays a role: The paper suggests that prosodic position (word-final vs. initial) also influences the voicing profile of [r], with devoicing being more pronounced in final positions even within similar phonological contexts.

Key Results

The study found that the Polish sonorant [r] devoices in word- and syllable-final positions, particularly when preceded by a voiceless stop. For instance, in the context of a voiceless stop followed by [r] in word- or syllable-final position, [r] was devoiced in over 80% of the exemplars. This devoicing tendency was observed to be stronger in final positions compared to initial positions, even when the preceding phonological context was the same. The standard deviation values for the voicing profiles of [r] indicated a relatively stable pattern within the analyzed contexts.

Limitations and Future Work

The paper acknowledges that while the automated tools provide a frame-by-frame analysis, the "voicing decision" is often categorical (0 or 1). The interpolation used to obtain values between these states is a simplification. The study primarily focuses on the sonorant [r], with other sonorants being "categorically voiced." Future work could explore the nuances of voicing in these other sonorants and investigate a wider range of phonetic and prosodic contexts. The authors also mention that dialectal variations and socio-phonetic features can influence sonorant devoicing, which was not deeply explored in this specific study.

Summary updated

BibTeX
@inproceedings{sieczkowska2009voicing,
  author = {J. Sieczkowska and Bernd Möbius and Antje Schweitzer and Michael Walsh and Grzegorz Dogil},
  title = {{Voicing profile of Polish sonorants: [r] in obstruent clusters}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure