TL;DR
This paper investigates how acoustic features like pitch and speaking rate signal formality in Japanese conversations. By analyzing natural speech, the authors found that informal Japanese speech is characterized by faster articulation, a wider pitch range, and a higher average pitch, suggesting these acoustic cues play a significant role in conveying social context.
What Problem Does This Paper Address?
Understanding how speakers convey social information, such as formality, through their speech is a crucial area in linguistics and speech technology. While Japanese has well-established grammatical and lexical markers of formality (like honorifics), the acoustic correlates of these distinctions in natural, spontaneous speech are less understood. Previous research often relied on lab-elicited speech, which may not reflect real-world conversational dynamics. This paper aims to bridge this gap by examining acoustic properties in naturally occurring Japanese conversations, seeking to identify reliable cues that signal formality and could potentially be used in speech synthesis or recognition systems.
Methods
The study employed a multi-stage approach:
- Pilot Study: A controlled experiment was conducted to identify potential acoustic correlates of informal speech. This involved analyzing speech containing specific informal phonetic phenomena (geminate contractions) and comparing them to their non-geminate counterparts. Key variables examined were mean fundamental frequency (f0), articulation rate, and amplitude.
- Data Collection: A new corpus of conversational Japanese was created. One-on-one interviews were conducted in a relaxed setting to encourage natural speech. The interviews were designed to gradually shift from formal to informal registers.
- Data Annotation: Utterances were meticulously labeled as formal, informal, or read speech based on a defined set of criteria, including verb forms, sentence-final particles, and specific lexical items. Ambiguous cases were resolved with the help of a native speaker.
- Acoustic Analysis: Standard acoustic analysis tools (Praat) were used to extract features like mean f0 and articulation rate. For a more in-depth analysis of pitch variation, a functional data analysis (FDA) approach was adopted. This involved fitting cubic polynomials to normalized pitch contours and analyzing the coefficients of these polynomials, which represent different aspects of pitch variation (e.g., overall slope, curvature).
Datasets
- Pilot Study Data: Speech from 5 native Japanese speakers (2 female, 3 male, aged 23-31) producing carrier sentences with and without geminate contractions.
- Main Study Corpus: Approximately 5 hours of recorded conversational Japanese from 10 native speakers (5 male, 5 female, aged 31-45) collected via one-on-one interviews. This corpus was annotated into 2,697 utterances (416 formal, 2,068 informal, 214 read).
Key Insights and Ideas
- Acoustic Cues Signal Formality: The study provides strong evidence that acoustic features are not merely byproducts of speech but actively contribute to signaling formality in Japanese.
- Informal Speech is More Dynamic: Informal Japanese speech exhibits greater variability and flexibility in its acoustic realization, particularly in terms of articulation rate and pitch range.
- Pitch Range as a Key Indicator: A wider pitch range appears to be a significant characteristic of informal speech, suggesting speakers use more expressive pitch contours.
- Functional Data Analysis for Pitch: The application of FDA proved valuable for analyzing pitch contours, overcoming limitations of traditional methods like pitch doubling errors and providing a more nuanced understanding of pitch variation.
Key Results
- Articulation Rate: Informal speech had a significantly higher articulation rate (7.82 moras/second) compared to formal speech (6.64 moras/second) and read speech (6.93 moras/second).
- Mean Fundamental Frequency (f0): Informal speech exhibited a significantly higher mean f0 (178.7 Hz) than formal speech (158.6 Hz) and read speech (159.7 Hz).
- f0 Range: Informal speech showed a significantly larger mean f0 range (228.72 Hz) compared to formal speech (178.44 Hz) and read speech (172.69 Hz).
- FDA Coefficients: Binary logistic regression analysis indicated that coefficients related to curvature, slope, and average height of the pitch contour were significant predictors of formality, with informal speech generally showing higher values for these coefficients.
Limitations and Future Work
- Subjectivity in Annotation: The authors acknowledge that judging formality is inherently subjective, and while consistent criteria were used, some ambiguity in labeling is unavoidable.
- Sample Size and Diversity: The corpus, while new, is relatively small. Future work could involve larger and more diverse corpora to increase the generalizability of the findings.
- Model Refinement: While the study shows potential for acoustic cues to predict formality, further refinement of predictive models is necessary.
- Interaction with Other Cues: Future research could explore how these acoustic cues interact with lexical and grammatical formality markers, and how they are perceived by listeners.
- Cross-linguistic Comparisons: Extending this analysis to other languages would provide valuable insights into the universality and language-specificity of acoustic formality cues.
Summary updated