Introduction
In the field of Text-to-Speech (TTS) synthesis, achieving natural-sounding prosody—the rhythm, stress, and intonation of speech—remains a significant challenge. While statement-like utterances often follow predictable patterns, interrogative sentences (questions) exhibit complex prosodic structures, particularly in tonal languages like Mandarin Chinese. Current systems frequently struggle with the "prosody gap," where synthesized questions sound robotic due to incorrect stress assignment. This paper proposes a computational model that leverages linguistic syntax to predict stress locations in Mandarin questions, thereby enhancing the naturalness of spoken dialog systems.
Methodology: Syntax-Stress Mapping
The core of the proposed approach is a "Syntax-Stress Mapping" strategy that translates grammatical structures into prosodic emphasis.
Dependency Parsing
The researchers utilize the Stanford Parser to generate dependency triplets for each sentence. These triplets define the grammatical relationships between words, such as Subject-Predicate (nsubj) or Verb-Object (dobj).
Rule-Based Mapping and Scoring
By conducting perceptual experiments, the authors established stable stress patterns for various dependency relations. For instance:
- Subject-Predicate (
nsubj): The subject (the first word) typically receives stress. - Verb-Object (
dobj): The object (the second word) usually carries the stress.
Beyond simple relations, the model incorporates syntax depth. The hypothesis is that stress "transmits" down the syntax tree; words located deeper in the hierarchy are more likely to be the "nuclear stress" of the sentence. A final stress score is calculated for each word by summing the weighted scores of all its associated dependency relations.
Binarization
To determine the final stress assignment, the model employs a binarization threshold. The top 20% of words by score are designated as "stressed," mirroring the natural distribution of prominence found in human speech.
Experimental Setup and Results
The study utilized a custom corpus of 400 Mandarin interrogative sentences ending in question particles (e.g., ma, ne). These were recorded by a professional male speaker and manually labeled for prominence by experts to create a ground-truth dataset.
The model achieved a weighted accuracy of 74% in mapping syntax to stress. Human evaluation further validated these results, with acceptance rates ranging from 76% to 80% among both expert and naive listeners. Interestingly, the model's performance was significantly higher for shorter sentences (78% acceptance for <8 words) compared to longer, more complex ones (54% acceptance for >10 words), indicating that increased grammatical complexity introduces more "noise" into the syntax-prosody relationship.
Discussion and Future Work
While the results are promising, several challenges remain. The model's reliance on the Stanford Parser means that any parsing errors directly impact stress assignment. Additionally, in Mandarin, the inherent lexical tone of a word can sometimes interfere with the perception of stress. Finally, the current model operates on isolated sentences and does not yet account for discourse context, which often shifts stress based on previous conversational turns.
Future research will focus on integrating this mapping into real-time TTS engines and exploring ways to incorporate broader contextual information to further refine prosodic naturalness.
Summary updated