Specification in context - devoicing processes in Polish, French, american English and German sonorants

Jagoda Sieczkowska, Bernd Möbius, Grzegorz DogilInterspeech 2010

TL;DR

This paper investigates how sonorants (like 'l', 'r', 'm', 'n') change their voicing (vocal fold vibration) depending on their linguistic context in Polish, French, American English, and German. They found that while Polish and French sonorants are primarily licensed for voicing by syllable structure, Germanic languages like English and German are more influenced by the "breathiness" (spread glottis) of preceding consonants.

What Problem Does This Paper Address?

Voicing is a fundamental phonetic feature distinguishing many consonants (e.g., 'p' vs. 'b'). While it's well-studied for stops, sonorants are generally assumed to be voiced. However, this paper explores how sonorants can devoice under specific phonetic and phonological conditions across different languages. The authors aim to understand the temporal dynamics of this devoicing and identify the linguistic factors (like surrounding sounds, syllable structure) that trigger it, particularly focusing on the rhotic sounds ('r'). This is important for building more accurate speech synthesis and recognition systems that account for these subtle, context-dependent variations.

Methods

The core of the methodology involves automatic voicing extraction from speech databases. They used a frame-by-frame analysis of fundamental frequency (F0) to determine voicing probabilities for sonorants. This was done at specific time points (10%, 20%, 90%) within the sonorant's duration. Crucially, they also extracted contextual information, including the sonorant's phonetic properties, its position within the word and syllable, and its phonotactic environment (i.e., the sounds immediately surrounding it). The results were then averaged and visualized using R.

Datasets

The study utilized four speech databases:

  • Polish Speech Corpus: Recorded by a professional male speaker for unit selection speech synthesis.
  • French Speech Corpus: Based on data from SVOX AG, including phone durations, syllabic/word/phrase positions, and binary voicing decisions.
  • Boston University Radio News Corpus: Recordings of FM radio news announcers for American English, specifically focusing on the FIA speaker's lab news materials.
  • MS Corpus: A large German speech corpus designed for unit selection speech synthesis (SmartKom project), containing extensive phonetic and contextual information.

Key Insights and Ideas

  • Sonorants aren't always voiced: Contrary to a common assumption, sonorants can and do devoice, especially in specific contexts.
  • Language-specific licensing: The "license" for a sonorant to be voiced differs across languages. Polish and French rely heavily on syllable structure, while English and German are more influenced by the "spread glottis" feature of preceding voiceless obstruents.
  • Rhotics are key: The 'r' sounds (rhotics) showed the most significant variations in voicing, making them a crucial focus for understanding these processes.
  • Desyllabification matters: In Polish and French, sonorants in certain word-final clusters with voiceless obstruents can be "desyllabified," leading to devoicing. This is linked to the Sonority Sequencing Principle.
  • "Spread Glottis" contagion: In Germanic languages (English and German), the feature of breathiness ([spread glottis]) from a preceding voiceless obstruent can transfer to the following sonorant, causing it to devoice.
  • Syllable boundaries influence German devoicing: For German sonorants in obstruent-sonorant clusters, the presence or absence of a syllable boundary between them significantly impacts the extent of devoicing.

Key Results

  • Polish: Sonorants devoice when between voiceless consonants or after a voiceless obstruent before a pause, often due to desyllabification.
  • French: Similar to Polish, the rhotic [R] shows significant devoicing in word-final positions after a voiceless obstruent.
  • American English & German: Sonorants in obstruent-sonorant clusters, particularly in word-initial and medial positions, exhibit devoicing. This is attributed to the transfer of the [spread glottis] feature from the preceding obstruent. German [R] showed the lowest voicing probabilities in these contexts.
  • Temporal profiles: The study provides detailed temporal profiles of voicing probabilities for sonorants across different languages and contexts, illustrating the gradual nature of devoicing.

Limitations and Future Work

  • Focus on Rhotics: While the study investigated all sonorants, the most detailed analysis and discussion focused on rhotics due to their greater voicing variability.
  • Automatic vs. Manual Annotation: The study mentions that some databases were automatically labeled and then manually corrected, while others were only automatically labeled. This could introduce varying levels of annotation accuracy.
  • Future Work: The authors suggest that further investigation into the interplay of phonological and phonetic factors across more languages would be beneficial. The findings also have direct implications for improving speech synthesis and recognition systems by incorporating these context-dependent voicing patterns.

Summary updated

BibTeX
@inproceedings{sieczkowska2010specification,
  author = {Jagoda Sieczkowska and Bernd Möbius and Grzegorz Dogil},
  title = {{Specification in context - devoicing processes in Polish, French, american English and German sonorants}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure