TL;DR
This paper challenges the assumption that formant targets (the ideal positions of vocal tract resonances) are style-independent. By modeling clear and conversational speech separately, they demonstrate that speakers adjust their target formant positions based on speaking style, leading to improved vowel classification accuracy when these style-dependent targets are used.
What Problem Does This Paper Address?
Speech intelligibility is often enhanced in "clear" speech compared to "conversational" speech. While acoustic differences are known, the underlying articulatory mechanisms are less understood. Specifically, this paper investigates whether the intended target positions of vocal tract resonances (formants) for speech sounds change depending on whether a speaker is producing clear or conversational speech. Previous work often assumed these targets were fixed, regardless of speaking style. This assumption might lead to inaccuracies in speech modeling and recognition systems, especially when dealing with different speaking styles.
Methods
The authors propose and evaluate a formant contour model that describes how formant frequencies change over time within a CVC (consonant-vowel-consonant) word. The core of their approach is to estimate "formant targets" for consonants and vowels. They compare two methods:
- Style-Independent Target Estimation: A single set of target values is estimated for each phoneme, assumed to be the same for both clear and conversational speech.
- Style-Dependent Target Estimation: Separate target values are estimated for each phoneme for clear speech and for conversational speech.
They use a formant contour model equation that combines target values with coarticulation functions (sigmoid functions that describe the influence of neighboring sounds). Model parameters are estimated by minimizing an error function that measures the difference between the modeled and observed formant contours. This estimation is performed using a steepest ascent hill-climbing optimization approach.
Datasets
The study uses a corpus of 242 meaningful English CVC words spoken by a single male speaker in two styles: clear (CLR) and conversational (CNV). Each word was spoken twice in each style, resulting in a total of 968 tokens. The data is split into training and testing sets using a jackknife procedure for evaluation.
Key Insights and Ideas
- Formant targets are not style-invariant: The most significant finding is that speakers adjust their intended formant target positions based on speaking style. This is supported by the improved model fitting and better clustering of estimated targets when style-dependent models are used.
- Style-dependent models improve fit: The mean error rate (a measure of how well the model fits the data) is significantly reduced when using style-dependent target estimation compared to style-independent estimation.
- Consonant targets show style differences: Particularly for the consonant /t/, the estimated target values show distinct clusters for CNV and CLR speech, providing strong evidence for style dependency.
- Vowel classification benefits from style-aware targets: When using the estimated formant targets to classify vowels, the style-independent token-based targets significantly improve accuracy for conversational speech compared to using observed steady-state formant values. This suggests that these model-derived targets capture more robust information about the vowel, independent of its context and speaking style.
Key Results
- Reduced Error Rate: The mean error rate (Es,target) in Bark was reduced with style-dependent target estimation for both training and test sets. For example, the test set error for style-independent targets was around 0.1164-0.1572 Bark, while style-dependent targets achieved 0.1091-0.1362 Bark.
- Improved Vowel Classification: Token-based style-independent target classification achieved 82.19% accuracy for conversational speech, a significant improvement over classification based on observed steady-state values (73.97%). For clear speech, the improvement was from 76.47% to 88.43% with style-independent targets and 76.05% to 88.23% with style-dependent targets.
- Better Clustering: Tables and figures demonstrate that style-dependent estimation leads to tighter clusters of estimated target values for both vowels and consonants, indicating more consistent and distinct targets per style.
Limitations and Future Work
- Single Speaker: The study was conducted with only one male speaker. The generalizability of these findings to other speakers, genders, and accents is not yet established.
- Limited Phonetic Context: While the model accounts for coarticulation, the initial experiments focused on limited CVC word contexts. Future work could explore more complex phonetic environments.
- Future Directions: The authors suggest that the improved vowel classification using token-based targets, even with style-independent global targets, shows promise for automatic speech recognition (ASR) systems. This approach could potentially improve ASR performance by providing more robust representations of speech sounds, especially in noisy or varied speaking conditions. Further research could explore incorporating explicit style detection or adaptation into ASR models.
Summary updated