TL;DR
This paper introduces a novel methodology for automatically detecting prominent syllables in spoken French, a crucial but labor-intensive task in prosodic analysis. By combining acoustic parameters with a rigorously annotated corpus, they achieve an impressive 84.1% agreement with human annotators, paving the way for more efficient and scalable prosodic research.
What Problem Does This Paper Address?
Analyzing large spoken language corpora often requires detailed prosodic annotation, with the identification of "prominent" syllables being a cornerstone. Prominence is what makes a syllable stand out perceptually, forming the backbone of prosodic structure. However, manually annotating this is incredibly time-consuming and, crucially, suffers from significant inter-annotator variability. Even expert phoneticians disagree on what constitutes a prominent syllable, with one study showing a variation from 19% to 49% of syllables being marked as prominent in the same recording. This inconsistency hinders reliable data exchange and large-scale analysis. Furthermore, existing automatic methods often struggle with the complexity of acoustic correlates and the need for robust, theoretically independent approaches.
Methods
The authors propose a two-step methodology for automatic prominence detection.
- Nuclei Segmentation and Pitch Stylization: This step leverages an adapted version of the "Prosogram" tool. It focuses on identifying vocalic nuclei within syllables and then stylizing their pitch contours. Modifications were made to Prosogram to improve nucleus detection, even in challenging cases, ensuring a higher coverage of syllables.
- Acoustic Parameter Computation: For each identified nucleus, two key acoustic prosodic parameters are computed:
- Syllable Duration: Measured in milliseconds, this is preferred over nucleus duration to avoid dependency on consonant onsets/codas.
- Maximum Pitch of the Nucleus: Measured in semitones.
Crucially, these parameters are not used in absolute terms but are relativized to adjacent syllables. This "local rate" and "local pitch" are calculated using a weighted average of preceding and succeeding syllables. The process also incorporates constraints from pauses and a "delivery tier" (discussed below) to avoid using unreliable adjacent syllables in the calculation.
Datasets
The study utilizes a newly compiled 18-minute corpus of spontaneous spoken French. This corpus consists of:
- "Itineraries": 4 recordings (2 male, 2 female speakers) where people were asked for directions. Durations range from 46 to 203 seconds.
- "Radio Interviews": 2 recordings (1 male, 1 female), each 5 minutes long.
This corpus was manually annotated using a strict methodology, which is a key contribution of the paper. The manual annotation process involved two annotators independently labeling syllables for both prominence and "speech delivery" issues (like hesitations, false starts, etc.). Disagreements were resolved through discussion. This rigorous annotation process is essential for training and evaluating the automatic detection system.
Key Insights and Ideas
- The Importance of Relativization: Using acoustic parameters relative to neighboring syllables, rather than absolute values, significantly improves robustness and independence from speaker and speech rate variations.
- Excluding "Delivery Tier" Data is Crucial: The paper strongly emphasizes the need to manually identify and exclude "non-valid" syllables (hesitations, interruptions, noise, etc.) from the training and evaluation data. Including these significantly degrades the performance of the automatic system.
- Prominence is Perceptual, Not Just Lexical: The methodology focuses on a neutral phonetic definition of prominence as perceptual salience, independent of theoretical frameworks or lexical stress, making it more broadly applicable.
- Correlation Between Acoustic Parameters: Figure 1 visually demonstrates a clear correlation between relative pitch and relative duration for prominent vs. non-prominent syllables, suggesting these two parameters are indeed strong indicators.
Key Results
The core quantitative result is the 84.1% agreement achieved by the automatic prominence detection system on a discriminant analysis of valid syllables. This figure is remarkably consistent across various testing scenarios:
- Closed-test: 84.1% overall (best sub-corpus 87.8%, poorest 79.7%).
- Cross-type test (radio vs. itinerary): 83.9% average.
- Cross-gender test (male vs. female): 84.1%.
These results indicate that the chosen acoustic parameters are largely independent of corpus type, speaker gender, and speech rate. The exclusion of "junk" syllables (marked with %) improved performance from 73.2% to 84.1%.
Limitations and Future Work
- Manual Annotation Burden: Despite the automation, the initial manual annotation process is still very time-consuming, requiring annotators to spend 4-5 times the duration of the corpus for their own labeling, and a similar amount for agreement sessions.
- Future Work: The authors suggest several avenues for future research:
- Utilizing enhanced plotting tools (like Prosogram's) to diagnose errors and refine the system.
- Optimizing the various parameters of their detection tool.
- Testing other acoustic parameters and applying the methodology to different corpora.
- Categorizing different types of prominences.
- Adapting the methodology and tools to other languages.
Summary updated