Acoustic analysis of Czech stress: intonation, duration and intensity revisited

Tomás Dubeda, Jan VotrubecInterspeech 2005

TL;DR

This paper investigates the acoustic correlates of stress in Czech, a language often described as having weak and subtle stress. Using a neural network to predict accent placement without lexical information, the study finds that fundamental frequency (pitch) is the most significant predictor, suggesting that Czech stress, while subtle, has objective acoustic underpinnings.

What Problem Does This Paper Address?

The existence and acoustic realization of stress in Czech have been a subject of debate. While Czech is generally described as having word-initial stress, many researchers note that this stress is often weak and its acoustic correlates (intonation, duration, intensity) are subtle and sometimes inhibited by other phonological rules (like vowel length). This raises the question of whether perceived accents in Czech have a truly objective, acoustically measurable basis or if they are more a matter of listener perception influenced by other factors. The paper aims to provide empirical evidence by analyzing acoustic features and their predictive power for accent placement.

Methods

The researchers employed a neural network approach to predict accent placement. They used a back-propagation architecture with an input layer representing prosodic features of the current and surrounding syllables, a hidden layer, and an output neuron indicating the probability of accent.

Key methodological aspects include:

  • Prosodic Feature Extraction: Intonation (fundamental frequency, F0), duration, and intensity were extracted for each syllable.
  • Normalization: Raw prosodic data was normalized to account for intrinsic properties of segments and syllable structure, aiming to isolate stress-related variations. F0 was measured at three points within the syllable (20%, 50%, 80%) to capture finer contour details.
  • Neural Network Training: The network was trained to predict accent presence (1) or absence (0) based on prosodic descriptions of the target syllable and its immediate neighbors (preceding and following).
  • Parameter Combinations: The study systematically tested the predictive power of different combinations of prosodic parameters (F0, duration, intensity) to understand their individual and combined roles.
  • Stochastic Approach: The use of a neural network and multiple runs with random initialization aimed to capture stochastic regularities and provide a robust assessment.

Datasets

The study utilized a single dataset:

  • Informative Text Corpus: A 970-syllable corpus of informative text read aloud by a male speaker. This corpus was manually segmented and accented by a human listener. 32% of the syllables were marked as accented. Pauses were also treated as syllables with specific parameters.

Key Insights and Ideas

  • Objective Basis for Czech Stress: The study provides strong evidence that Czech stress, despite its subtlety, has objective acoustic correlates that can be reliably predicted. The "no-accent hypothesis" for Czech is rejected.
  • Dominance of Fundamental Frequency (F0): Pitch (F0) emerged as the most powerful predictor of stress, both when used alone and in combination with other parameters. This suggests that intonational contours are crucial for signaling stress in Czech.
  • Limited Role of Duration and Intensity: Duration and intensity played a less significant role in prediction compared to F0. This aligns with previous descriptions of Czech stress, where duration is constrained by phonological length and intensity is not a strong correlate.
  • Importance of Context: The prosodic description of surrounding syllables (context) was essential for accurate prediction, highlighting the importance of stress units and prosodic discontinuities in the Czech language.
  • Neural Network as a Perceptual Model: The neural network's ability to predict accent suggests it can capture regularities similar to human perception, even without explicit lexical information.

Key Results

  • Overall Prediction Score: The neural network achieved a mean prediction score of 80% when using all three prosodic parameters (F0, duration, intensity) and context. This is a significant improvement over chance (which would be around 68% due to the default unstressed setting).
  • F0 Alone: F0 alone achieved a prediction score of 69%, demonstrating its strong predictive power.
  • Best Performance: The highest prediction score (80%) was achieved when all three parameters (F0, duration, intensity) were used as input.
  • Pauses: Pauses were predicted with a mean score of 80%, comparable to the overall mean for syllables.
  • Error Analysis: The analysis of prediction errors revealed that accent omissions (network missed an accent) and additions (network falsely predicted an accent) were distributed, with some errors potentially linked to normalization issues or the inherent weakness of some acoustic cues.

Limitations and Future Work

  • Normalization Challenges: While normalization was elaborate, the authors acknowledge potential normalization problems, particularly for duration and intensity, as suggested by some error patterns.
  • Network vs. Human Perception: The paper notes that the network's independent prediction for each syllable doesn't fully replicate human perception, which likely integrates information more dynamically.
  • Limited Corpus: The study used a single speaker and a specific type of text (informative). Future work could explore different text genres, speakers, and comparative studies with languages having different stress systems.
  • Technical Optimization: The authors suggest further research into technical optimization of the network.
  • Rhythmic Prediction: The authors hypothesize that rhythmic prediction might help improve accuracy, especially for accent omissions.

Summary updated

BibTeX
@inproceedings{dubeda2005acoustic,
  author = {Tomás Dubeda and Jan Votrubec},
  title = {{Acoustic analysis of Czech stress: intonation, duration and intensity revisited}},
  booktitle = {Interspeech},
  year = {2005},
}
Figure