Analysis of factors involved in the choice of rising or non-rising intonation in question utterances appearing in conversational speech

Hiroaki Hatano, Miyako Kiso, Carlos T. IshiInterspeech 2013

TL;DR

This paper investigates why Japanese questions sometimes end with a rising intonation and sometimes with a falling or flat one. They found that the choice is influenced by the type of question, the specific grammatical ending, and the relationship between speakers, suggesting intonation is a nuanced signal beyond just asking a question.

What Problem Does This Paper Address?

In many languages, questions are signaled by a rising intonation at the end of a sentence. However, this isn't always the case, especially in spontaneous, natural speech. This paper focuses on Japanese conversational speech, where questions frequently end with non-rising intonation. The authors aim to uncover the linguistic and social factors that determine whether a Japanese question utterance will have a rising or non-rising intonation, a phenomenon that has not been fully clarified by previous research, which often relied on written data or introspection. Understanding these factors is crucial for improving the naturalness of speech synthesis and the accuracy of dialogue systems.

Methods

The researchers analyzed a corpus of Japanese conversational speech, segmenting utterances into phrases and classifying them by dialogue act, specifically focusing on "question utterances." They developed a sophisticated method for classifying phrase-final tones based on fundamental frequency (FO) movements and pitch resets. Key parameters included:

  • F0 movement (Fomove): Quantifying the rise or fall of pitch within the final syllable of a phrase.
  • Pitch reset (Foreset): Measuring the pitch change between the final syllable of a phrase and the preceding syllable.
  • Duration: Categorizing phrase finals as short or long.

These parameters were combined to define three main tone categories: rise (Rs), fall (Fa), and flat (Ft), further distinguished by the presence or absence of a pitch reset (Rt). Additionally, they classified question types based on semantic categories and analyzed the grammatical morphemes at the end of utterances. They also considered interpersonal relationships, categorizing interactions by familiarity, social status, and a "concern" factor.

Datasets

The study utilized a Japanese conversational database comprising 58 dialogue sessions from 28 speakers, totaling approximately 600 minutes of free conversation. This data included simultaneous recordings of speech and electroglottograph (EGG) signals. A total of 3,661 "question utterances" were selected for analysis after annotation by multiple annotators.

Key Insights and Ideas

  • Beyond Yes/No: While rising intonation is common for questions, the study reveals that non-rising intonation is prevalent in specific question types like "request for agreement," "open-type questions," and "self-directed doubt." This suggests that the speaker's assumption about whether the interlocutor possesses the answer significantly influences intonation.
  • Grammar Matters: The grammatical ending of a question plays a crucial role. Utterances ending with specific question-related final particles (like /yone/, /ne/, /na/, /kana/) are more likely to have non-rising intonation, possibly because these particles already convey attitudinal meanings or agreement.
  • Social Cues: Interpersonal relationships, particularly the "concern" factor (indicating care towards the interlocutor due to age, status, or unfamiliarity), significantly impact intonation. Speakers tend to use less rising intonation when expressing concern, perhaps to avoid being perceived as overly intrusive.
  • Nuanced Communication: The findings highlight that intonation in Japanese questions is not a simple binary switch but a complex interplay of linguistic content, grammatical form, and social context.

Key Results

  • Approximately 20% of question utterances in the dataset exhibited neither rising intonation nor a pitch reset.
  • Over 50% of "request for agreement," "open-type questions," "backchannel-type questions," and "self-directed doubt-type questions" featured non-rising intonation.
  • Utterances ending with question-related final particles showed non-rising intonation in over 50% of cases, while those ending with other morphemes had rising intonation or a pitch reset in over 80% of cases.
  • One-way ANOVA tests confirmed that question type, phrase final morphemes, and the "concern" factor all significantly influenced pitch movement (Fomove).

Limitations and Future Work

The authors acknowledge that categories with fewer than 30 utterances were excluded from some analyses, which might limit the generalizability of findings for those specific types. They also note that their analysis focused on Japanese conversational speech and suggest that similar strategies could be applied to extend the analysis to other languages. Future work could explore the interaction between these factors more deeply and investigate how these intonational patterns are perceived by native speakers.

Summary updated

BibTeX
@inproceedings{hatano2013analysis,
  author = {Hiroaki Hatano and Miyako Kiso and Carlos T. Ishi},
  title = {{Analysis of factors involved in the choice of rising or non-rising intonation in question utterances appearing in conversational speech}},
  booktitle = {Interspeech},
  year = {2013},
}
Figure