Advances in regional accent clustering in Swedish

Giampiero SalviInterspeech 2005

TL;DR

This paper introduces a novel approach to analyzing regional pronunciation variations in Swedish by applying advanced data mining techniques, specifically EM algorithm and hierarchical clustering, to a large speech database. This method allows for a more nuanced understanding of accent differences by considering cross-phoneme similarities and providing interpretable insights into the acoustic features driving these variations.

What Problem Does This Paper Address?

Analyzing regional pronunciation variations in language is challenging due to the sheer volume of data and the complexity of phonetic transcription. Traditional statistical methods struggle with large datasets, while data-mining techniques, though powerful for classification, often lack interpretability. This paper aims to bridge this gap by using data mining to uncover and understand regional accent differences in Swedish, moving beyond isolated phoneme analysis to explore how acoustic properties of different sounds interact across regions.

Methods

The core of the methodology involves using the Expectation-Maximization (EM) algorithm within a hidden Markov model (HMM) framework to estimate phonetic model parameters from a large, partially labeled speech database. This circumvents the need for exhaustive manual phonetic transcriptions. These HMMs, representing the pronunciation of phoneme segments (initial, middle, final) within specific regions, are then clustered using agglomerative hierarchical clustering. The dissimilarity between these models is measured using the Bhattacharyya distance. To interpret the clustering results, discriminant analysis is employed to identify which acoustic parameters (Mel-frequency cepstral coefficients and their derivatives) are most influential in distinguishing between clusters.

Datasets

The study utilizes the Swedish SpeechDat FDB5000 telephone speech database. This database contains utterances from 5000 speakers, recorded over the fixed telephone network, with lexical-level labeling and a lexicon of 46 phonetic symbols. For this study, accent information was the primary focus. The official SpeechDat training set was used, comprising 4500 speakers and 270 hours of speech, resulting in approximately 97 million 10-millisecond frames.

Key Insights and Ideas

  • Cross-Phoneme Similarity: A significant advancement is the ability to cluster phoneme segments across different phonemes, revealing that acoustic similarities can lead to unexpected groupings, rather than just regional variations of the same phoneme.
  • Interpretable Accent Features: Discriminant analysis provides a powerful tool to interpret the clustering results, highlighting which specific acoustic features (e.g., energy, spectral slope) are most responsible for differentiating regional accents.
  • Hierarchical Clustering for Interpretation: The use of hierarchical clustering is crucial for visualizing and understanding pronunciation variations at different levels of detail, allowing researchers to explore both broad trends and specific nuances.
  • Regional Variation as Context: Regional accent information is treated as a form of contextual information, similar to co-articulation effects, allowing for the modeling of pronunciation patterns specific to geographical areas.

Key Results

The study successfully clustered regional pronunciation variants of Swedish phonemes. The clustering process revealed that, generally, differences in phoneme segments (initial, middle, final) and phonemes themselves are distinguished before regional effects become prominent. However, notable exceptions highlight interesting accent-driven sound changes. For instance, the retroflex allophone of /s/ in southern Sweden clusters with the main pronunciation of /s/, and the vowel /ø/ in Gotland groups with variants of /u/ and /e:/ in southwestern Skåne. Discriminant analysis showed that initial acoustic features like energy and spectral slope were highly informative for distinguishing clusters.

Limitations and Future Work

The authors acknowledge that the models used are likely "under-dimensioned" given the vast amount of data, leading to monotonically increasing cluster validation indexes. This suggests that more complex models might be necessary for a complete representation. The paper suggests that future work could explore more sophisticated methods for quantifying the fit of clustering models and further investigate the specific acoustic phenomena revealed by the clustering tree, potentially leading to improved speech recognition systems that are more robust to regional accent variations.

Summary updated

BibTeX
@inproceedings{salvi2005advances,
  author = {Giampiero Salvi},
  title = {{Advances in regional accent clustering in Swedish}},
  booktitle = {Interspeech},
  year = {2005},
}
Figure