TL;DR
This paper investigates whether listeners can identify different speech genres (like political speeches or sports commentary) using only prosodic cues (intonation, rhythm, etc.) in delexicalized speech. They found that listeners can indeed distinguish genres, with performance varying by genre and significantly influenced by the listener's native language background, suggesting prosody plays a crucial role in genre perception.
What Problem Does This Paper Address?
Understanding and classifying different types of spoken language, known as discourse genres (DGs), is a fundamental challenge in linguistics and speech processing. While DGs are typically defined by their context, social function, and linguistic markers, this paper focuses on a less explored aspect: the role of prosody – the rhythm, stress, and intonation of speech – in distinguishing these genres. The authors aim to determine if prosodic patterns alone are sufficient for listeners to identify DGs, and how this perception is affected by factors like language background. This is particularly relevant for developing more natural and context-aware speech synthesis systems.
Methods
The core of this research is a perceptual experiment. They created a corpus of French speech samples from four distinct discourse genres: church sermons, political speeches, journalistic reports, and sports commentary. To isolate the effect of prosody, the speech samples were delexicalized (linguistic content removed) and then presented to listeners.
Key methodological steps include:
- Corpus Creation: A 4-hour multi-media corpus was designed, focusing on male speakers and specific discourse contexts within each genre to reduce intra-variability.
- Stimuli Preparation: Speech utterances were segmented based on prosodic periods and discursive considerations. Crucially, linguistic content was removed using a band-pass filter to isolate prosodic features. Noise and reverberation were also reduced.
- Perceptual Experiment: 70 participants, with varying levels of French language proficiency and speech expertise, were asked to identify the DG of 40 delexicalized speech samples. They could indicate total confidence, confusion between two genres, or total indecision.
- Analysis: Identification performance was measured using Cohen's Kappa statistic. Statistical analyses (ANOVA) were used to assess the impact of language background and DG. Multidimensional Scaling (MDS) and hierarchical clustering were employed to visualize and quantify perceptual similarity between genres based on confusion patterns.
Datasets
The study utilized a newly created 4-hour French multi-media corpus for the perceptual experiment. This corpus was specifically designed to cover four discourse genres:
- Mass: Christian sermons
- Political: New year's speeches
- Journal: Radio news reports
- Sport commentary: Soccer commentary
The corpus was enriched with linguistic annotations like orthographical transcription and phonemic alignment.
Key Insights and Ideas
- Prosody as a Genre Marker: The study strongly supports the hypothesis that prosodic patterns are significant, perceptually salient markers of discourse genres, independent of specific speakers.
- Language Matters, But It's Complicated: While listeners can identify genres from prosody, their performance is significantly influenced by their native language. This suggests that abstract prosodic representations of DGs are at least partially language-dependent, and potentially even culture-dependent.
- Genre-Specific Prosodic Signatures: Different genres exhibit varying degrees of prosodic distinctiveness. Sports commentary, for instance, has a highly recognizable prosodic signature that is relatively consistent across language backgrounds, likely due to its imitative and expressive nature. Political speeches and church sermons, conversely, show greater dependence on language and cultural context for their prosodic identification.
- Perceptual Similarity Mirrors Conceptual Similarity: The clustering of genres based on listener confusion patterns closely aligns with conceptual classifications of these genres, suggesting that our intuitive understanding of genre is reflected in how we perceive their prosodic cues.
Key Results
- Overall Identification Performance: Listeners achieved a "fair" identification performance with an overall Cohen's Kappa of 0.45.
- Language Background Impact: Native French speakers performed significantly better than non-native French speakers, who in turn performed better than non-French speakers.
- Genre-Specific Performance: Identification was highest for sports commentary (Kappa=0.7), followed by journalistic discourse (Kappa=0.54), and then mass and political discourse (Kappa=0.38 and 0.34, respectively).
- Language-Genre Interaction: The influence of language background varied significantly across genres. For example, the language effect was strong for political discourse but less pronounced for sports commentary.
- Perceptual Clustering: MDS analysis revealed distinct clusters of genres based on perceptual confusion, with sports commentary standing apart from the others.
Limitations and Future Work
- Language Dependence: The study highlights that prosodic representations of DGs are at least partially dependent on language. The authors hypothesize this could extend to language classes or even culture, suggesting future work could explore cross-linguistic comparisons.
- Speech Synthesis Evaluation: The authors intend to use these results as a benchmark for evaluating the prosodic realism of speaking style speech synthesis systems.
- Further Discursive Effects: While prosody is a key factor, the authors note that other discursive effects likely contribute to the observed perceptual differences and suggest further investigation.
- Limited Genres and Speakers: The study focused on four specific genres and male speakers, leaving room for expansion to a wider range of genres and speaker demographics.
Summary updated