A corpus-based study of elderly and young speakers of European Portuguese: acoustic correlates and their impact on speech recognition performance

Thomas Pellegrini, Annika Hämäläinen, Philippe Boula de Mareüil, Michael Tjalve, Isabel Trancoso, Sara Candeias, Miguel Sales Dias, Daniela BragaInterspeech 2013

TL;DR

This paper investigates how the speech of elderly Portuguese speakers differs acoustically from younger speakers and how these differences impact automatic speech recognition (ASR) performance. They found that increased pausing and slower speech rates are key acoustic correlates of aging, significantly degrading ASR accuracy, and that younger speakers exhibit more vowel reduction.

What Problem Does This Paper Address?

As our population ages, there's a growing need for technologies that are accessible and usable by older adults, particularly speech interfaces. However, current ASR systems often perform poorly with elderly speech. This is partly because acoustic models are typically trained on younger speakers, and partly because aging itself can alter speech production. This paper aims to bridge this gap by systematically analyzing the acoustic differences between elderly and young European Portuguese speakers and, crucially, by linking these acoustic features to ASR performance. Understanding these age-related speech characteristics is vital for developing more robust and inclusive ASR systems.

Methods

The researchers employed a multi-faceted approach:

  • Acoustic Feature Extraction: They analyzed 14 acoustic features, including source-related features (pitch, jitter, shimmer, harmonics-to-noise ratio) and vowel-related features (formants, schwa realization, speech rate, pause duration, and phone duration). These were computed using the Praat software and an in-house ASR engine (Audimus) for forced alignment.
  • Decision Trees for Feature Importance: To identify the most discriminative features between age groups, they built decision trees. This method helps rank features by their relative importance in classifying speakers into young, senior (60-75), and elderly (75-90) categories.
  • Correlation with ASR Performance: They then investigated how these acoustic features correlated with Word Error Rate (WER) obtained from an ASR system. This step directly links acoustic changes to recognition accuracy.
  • Analysis of Vowel Reduction: A language-specific analysis focused on vowel reduction, particularly the schwa sound in European Portuguese, by comparing phone sequences used by the ASR system for frequent words.

Datasets

The study utilized two corpora of European Portuguese speech:

  • BD-PUBLICO: This corpus provided speech from young adult speakers (aged 19-30) reading newspaper sentences. It served as the baseline for young speaker data.
  • EASR: This corpus contained speech from elderly speakers (aged 60 and over) reading various prompts. For this study, only phonetically rich sentences were used to match the BD-PUBLICO material. The EASR speakers were further divided into two groups: 'S' (Senior, 60-75 years) and 'E' (Elderly, 75-90 years).

Key Insights and Ideas

  • Pauses are Key: The most significant acoustic differentiator between elderly and young speakers, regardless of gender, was the increased frequency and duration of pauses in elderly speech. This directly impacts speech rate and fluency.
  • Speech Rate Slows Down: As expected, speech rate (both word and phone level) generally decreased with age. This was particularly noticeable in female speakers.
  • Vowel Reduction Differences: A fascinating language-specific finding was that younger speakers exhibited more significant reduction of the schwa vowel compared to elderly speakers. This suggests a potential generational difference in articulation, not solely an aging effect.
  • Feature Importance Varies by Gender: While pauses were universally important, other features showed gender-specific relevance. For instance, word-level speech rate was a key differentiator for elderly females, whereas jitter was more critical for distinguishing between senior and elderly males.
  • ASR Impact is Clear: The study directly linked acoustic features to ASR performance, showing that increased pausing, slower speech rates, and a higher proportion of long phones were strongly correlated with higher word error rates.

Key Results

  • Increased Pausing: Speakers aged 60 and over showed a significantly higher percentage of pauses compared to young speakers.
  • Speech Rate Correlation: For female speakers, speech rate correlated with age. For male speakers, jitter distinguished between the 60-75 and older groups.
  • WER Correlation: Word error rate correlated most strongly with the use of pauses, speech rate, and the ratio of long phone realizations.
  • Vowel Reduction: Young adult speakers reduced schwas more than elderly speakers, with maintained schwas being 32% for young speakers versus 58-62% for older groups.
  • Decision Tree Outcomes: For females, word_sr was the most important feature after pauses, followed by shimmer and F1_[u]. For males, jitter was the primary differentiator between senior and elderly groups.

Limitations and Future Work

  • Generalizability: The study focused on European Portuguese. While the findings on pauses and speech rate are likely generalizable, language-specific features like vowel reduction might differ across languages.
  • Perception vs. Chronological Age: The study focused on chronological age. Future work should investigate how perceived age correlates with speech and ASR performance.
  • Other Voice Quality Features: The authors suggest exploring additional voice quality features like creakiness and breathiness.
  • Generation Gap vs. Aging: Further research is needed to disentangle whether the observed differences in vowel reduction are due to the natural aging process or a "generation gap" in speaking styles.
  • Perception Experiments: Conducting perception experiments to measure how human listeners perceive the age of speakers, in addition to chronological age, is recommended.

Summary updated

BibTeX
@inproceedings{pellegrini2013corpus,
  author = {Thomas Pellegrini and Annika Hämäläinen and Philippe Boula de Mareüil and Michael Tjalve and Isabel Trancoso and Sara Candeias and Miguel Sales Dias and Daniela Braga},
  title = {{A corpus-based study of elderly and young speakers of European Portuguese: acoustic correlates and their impact on speech recognition performance}},
  booktitle = {Interspeech},
  year = {2013},
}
Figure