TL;DR
This paper investigates the optimal level of linguistic analysis (word, syllable, or vowel) for detecting pitch accents in speech using only acoustic features. They found that a word-level analysis significantly outperforms syllable and vowel-level approaches, achieving 84.2% accuracy, suggesting that crucial accent information extends beyond the stressed syllable.
What Problem Does This Paper Address?
Pitch accents are crucial for conveying meaning in spoken language, signaling things like contrast, focus, and information status. While we intuitively understand their importance, automatically detecting them accurately is a challenging task. Previous research has explored various methods and levels of analysis, but it's been unclear whether focusing on individual vowels, syllables, or entire words yields the best results, especially when relying solely on acoustic cues. This ambiguity makes it difficult to compare different approaches and hinders the development of robust speech processing systems that rely on prosodic information. The authors aim to resolve this by directly comparing word, syllable, and vowel-level acoustic analysis for pitch accent detection on a common dataset using a consistent methodology.
Methods
The core of this paper is a comparative study of pitch accent detection at three different linguistic levels: word, syllable, and vowel. They employed logistic regression models trained on acoustic features extracted from these units. The acoustic features included pitch (f0), energy, and duration. To capture contextual information, they calculated the z-score of the maximum and mean pitch and energy over several "short range" (one previous/following region) and "long range" (two previous/following regions) contexts. For syllable and vowel analysis, these "long range" contexts were scaled to cover approximately three syllables or vowels, balancing the analysis window with the average number of syllables per word. To enable direct comparison, they also derived word-level predictions from syllable and vowel-level analyses by predicting a word as accented if any of its constituent syllables or vowels were predicted as accented.
Datasets
The experiments were conducted on the Boston University Radio News Corpus (BURNC). This corpus consists of 157.9 minutes of professionally read news from six speakers. It is prosodically annotated with full ToBI labeling, including pitch accent information at the syllable level. Forced alignment was used to generate time-aligned phone boundaries, which were then used to identify vowel regions. Syllable boundaries were generated by aligning forced-aligned phones with a syllabified lexicon.
Key Insights and Ideas
- Word-level analysis is king for acoustic pitch accent detection: The most striking finding is the consistent superiority of word-level analysis over syllable and vowel-level approaches. This suggests that the acoustic cues for pitch accent are not confined to the lexically stressed syllable but are distributed across the word.
- Context is crucial, but word-level analysis inherently captures more relevant context: While incorporating contextual acoustic features improved performance across all levels, the word-level analysis still outperformed others. This implies that the "context" relevant for accent detection is often best captured by considering surrounding words, not just surrounding syllables or vowels.
- Disparity between acoustic and communicative domains: The authors highlight that while acoustic properties are aligned with the lexically stressed syllable, the communicative impact of accenting (like contrast or focus) is often at the word level. This paper provides evidence that acoustic analysis should also align with this word-level communicative function for optimal detection.
- Limitations of syllable/vowel nuclei: The authors speculate that the narrow analysis windows of syllable and vowel nuclei might make their aggregated features more susceptible to noise and errors from forced alignment or syllabification.
Key Results
- The word-based approach achieved the highest accuracy at 84.2%.
- Syllable-based detection achieved 75.6% accuracy, and vowel-based detection achieved 68.5% accuracy (without contextual features).
- When contextual features were included, vowel-based classification showed a nearly 10% absolute increase in accuracy, reaching 77.4%. Syllable-based detection improved to 81.9%, and word-based detection improved to 84.2%.
- The F-measure for detecting accent was consistently higher for word-based results across all experiments.
Limitations and Future Work
The authors acknowledge that their word-based accuracy of 84.2% is lower than some previous work that incorporated lexical and syntactic information. However, their focus was on comparing acoustic analysis domains, and they explicitly excluded linguistic features to isolate the impact of the analysis level. They also note that their work relies on the accuracy of the BURNC annotation and the quality of forced alignment and syllabification.
For future work, they propose exploring techniques to transfer word-based predictions to syllables, which would allow for a more direct comparison with published syllable-based results. They also suggest that an automatic, acoustic syllable-nucleus detection approach might be able to generate more discriminative regions than the current forced-alignment and lexicon alignment techniques.
Summary updated