Accent assignment algorithm in Hungarian, based on syntactic analysis

Anne Tamm, Kálmán Abari, Gábor OlaszyInterspeech 2007

TL;DR

This paper tackles the challenge of automatically assigning word accents in Hungarian, a language with a complex, information-structure-driven prosody. They propose a hybrid approach combining a "shallow" statistical method with a "deep" syntactic analysis, finding that neither method alone is sufficient, but their combination yields improved results.

What Problem Does This Paper Address?

Accurate prosodic modeling is crucial for natural-sounding speech synthesis (TTS). A key component of prosody is word accent, which signals emphasis and can significantly alter the meaning or perceived naturalness of speech. Hungarian, unlike many fixed-word-order languages, has a flexible word order and a prosodic structure heavily influenced by information structure (like focus and topic). This makes traditional accent assignment methods, often tied to syntactic positions, less effective. The authors aim to develop an algorithm that can predict Hungarian word accents by leveraging both local acoustic cues and deeper syntactic information, addressing the limitations of existing approaches.

Methods

The paper presents two main methods for accent assignment:

  • Shallow Method: This approach relies on local, word-based, and short-distance statistical observations. It uses a set of rules and lists derived from acoustic perception and Fo (fundamental frequency) curve analysis. The core idea is to identify accent based on immediate word context and sentence structure.
  • Deep (Syntactic) Method: This method incorporates linguistic components to capture long-distance effects and phrase-internal accent distribution. It analyzes the sentence's syntactic structure, particularly the distinction between topic and predicate parts, and how focus and other information-structural elements influence accent. Key components include noun phrase (NP) analysis and morphological analysis.

The authors propose a hybrid approach, combining the strengths of both methods, as neither alone provides satisfactory results.

Datasets

The study uses a dataset of 1082 sentences from media texts, specifically news and weather forecasts. This corpus was used for both developing and evaluating the accent assignment algorithms. For evaluation, a subset of 52 sentences (756 words) was used for direct comparison between the shallow and deep methods.

Key Insights and Ideas

  • Hungarian Prosody is Information-Driven: The paper strongly emphasizes that Hungarian accent assignment is highly sensitive to information structural constraints (focus, topic), making it distinct from languages where accent is primarily determined by fixed syntactic positions.
  • Hybrid Approach is Necessary: Neither the shallow (local, statistical) nor the deep (syntactic, long-distance) method alone is sufficient for accurate accent assignment in Hungarian. Their complementary nature is key to improving performance.
  • Focus Has a Long-Distance Effect: The "deep" method highlights the significant impact of focus on accentuation, causing neutralization of accents in subsequent phrases. This is a crucial long-distance influence that shallow methods struggle to capture.
  • Error Complementarity: The mistakes made by the shallow and deep methods are often complementary, meaning that when one method errs, the other might be correct, further justifying their combination.
  • Error Types Matter: The authors distinguish between "overstress" (M1, assigning emphatic accent instead of neutral) and "understress" (M2, assigning neutral instead of emphatic/normal) errors. The deep method is better at reducing the more disturbing "overstress" errors.

Key Results

  • Shallow Method Performance: Achieved 12.8% word error rate (WER) on a test set, with 11.5% being M2 (understress) and 1.3% being M1 (overstress) errors.
  • Deep Method Performance: Achieved 11.9% WER, with 5% M2 and 85% M1 errors (note: the paper states M1 85 (87), M2 5 (10) in section 3.3, implying 85 M1 and 5 M2 errors for the deep method, and 87 M1 and 10 M2 for the shallow method, which is a bit confusingly presented. However, the key takeaway is the reduction in M2 errors).
  • Hybrid Advantage: While not explicitly quantified as a single combined score, the paper states that "mistakes are complementary" and the deep method's advantage lies in reducing the more disturbing M2 errors, implying that a combined system would outperform either individual method.

Limitations and Future Work

  • Dependency on Morpho-Syntactic Analysis: The deep (syntactic) method's accuracy is heavily dependent on the quality of the automatic morpho-syntactic analysis, which can introduce ambiguities (e.g., distinguishing between verb, noun, or adjective forms).
  • Ambiguity in Analysis: The deep method can lead to a higher number of ambiguities in the analysis, requiring further refinement.
  • Limited Extendibility of Shallow Method: The shallow method, while fast, is difficult to refine and expand without incorporating more linguistic analysis.
  • Future Work: The authors suggest that a more effective combination of the two approaches is needed to reduce the risk of amplifying errors from the morphological and syntactic analyzers. They also note that this approach could be applied to other non-configurational languages like Estonian.

Summary updated

BibTeX
@inproceedings{tamm2007accent,
  author = {Anne Tamm and Kálmán Abari and Gábor Olaszy},
  title = {{Accent assignment algorithm in Hungarian, based on syntactic analysis}},
  booktitle = {Interspeech},
  year = {2007},
}
Figure