Analysing the Performance of Automatic Speech Recognition for Ageing Voice: Does it Correlate with Dependency Level?

Frédéric Aman, Michel Vacher, Solange Rossato, François Portet2013

TL;DR

This paper investigates the performance of Automatic Speech Recognition (ASR) systems with aging voices, a crucial component for Ambient Assisted Living (AAL) technologies. They found that while chronological age is a factor, a person's level of functional dependency is a stronger predictor of ASR performance degradation, suggesting a more nuanced approach to AAL system design.

What Problem Does This Paper Address?

As global life expectancy increases, so does the population of elderly individuals. Ambient Assisted Living (AAL) technologies aim to help these individuals maintain independence and safety in their homes. Voice-based interfaces, particularly Automatic Speech Recognition (ASR), are a promising avenue for AAL due to their naturalness and accessibility, especially for those with physical or cognitive declines. However, existing ASR systems, often trained on younger, healthier voices, struggle with the atypical characteristics of aging speech (e.g., tremors, hesitations, altered articulation). While chronological age has been considered a primary factor in ASR performance degradation, this paper questions whether it's the sole or even the most significant predictor, especially in the context of AAL where functional abilities vary greatly among the elderly. The authors aim to identify which characteristics of aging voices challenge ASR and to explore if factors beyond chronological age, like functional dependency, better explain performance variations.

Methods

The researchers employed a multi-pronged approach:

  • Corpus Collection: They created a new French corpus, the "AD corpus," specifically for this study. This corpus includes distress and home automation commands, as well as casual sentences, recorded from both elderly (aged 62-94) and non-elderly (aged 18-64) speakers. They also utilized a separate aged corpus, "ERES38," for adaptation purposes.
  • Phoneme Analysis: To pinpoint specific speech challenges, they used forced alignment on the AD corpus. This technique maps spoken utterances to their phonetic representations, allowing them to compare alignment scores between aged and non-aged groups. Lower scores indicate a poorer match to the generic acoustic model, highlighting problematic phonemes.
  • Acoustic Adaptation: They applied Maximum Likelihood Linear Regression (MLLR) adaptation using the ERES38 corpus to create an acoustic model tailored for elderly voices. This is a standard technique for adapting ASR systems to new speaker characteristics.
  • Dependency Level Analysis: After adaptation, they analyzed the Word Error Rate (WER) of the elderly speakers in relation to their level of functional autonomy, as assessed by the French AGGIR grid. This grid categorizes individuals into six groups (GIR 1 to GIR 6) based on their loss of autonomy.

Datasets

  • AD Corpus (New): Collected by the authors, this corpus consists of distress, home automation, and casual sentences from 95 French speakers (43 elderly, 52 non-elderly). It was used for both evaluation and identifying phoneme challenges.
  • ERES38 Corpus: Acquired for model adaptation, this corpus contains speech from 22 elderly individuals.
  • BREF120 Corpus: Used for training the initial generic acoustic model, this is a large French speech corpus from 120 non-elderly speakers.

Key Insights and Ideas

  • Dependency Trumps Age: The most significant finding is that a speaker's level of functional dependency (as measured by the AGGIR grid) is a stronger predictor of ASR performance degradation than their chronological age alone. This suggests that the physical and cognitive impacts of aging, rather than just the passage of time, are the primary drivers of ASR challenges.
  • Specific Phoneme Vulnerabilities: The forced alignment analysis revealed that certain phoneme categories, particularly nasal vowels and plosive consonants, are more problematic for ASR systems when spoken by elderly individuals. This provides concrete targets for future acoustic model development.
  • Adaptation is Effective, but Nuance is Needed: Acoustic adaptation (MLLR) significantly reduces WER for elderly speakers, bringing performance closer to that of non-elderly speakers. However, the remaining variability highlights the need for more sophisticated adaptation strategies or models that account for individual differences beyond general age-related characteristics.
  • Common Characteristics in Aging Voices: The success of MLLR adaptation, even with a relatively small adaptation corpus (ERES38), suggests that aging voices share common characteristics that can be effectively modeled.

Key Results

  • Significant WER Increase: Without adaptation, the average WER for the elderly group was 43.47%, a substantial increase compared to the non-elderly group's 9.07% (a difference of 34.40%).
  • Adaptation Reduces WER: MLLR adaptation reduced the average WER for the elderly group to 14.52%, a significant improvement.
  • Dependency Levels Show Clear Differences:
    • GIR 2-3 (higher dependency): Average WER of 25.2%
    • GIR 4-5 (moderate dependency): Average WER of 13.2%
    • GIR 6 (low dependency/high autonomy): Average WER of 12.2%
    • Statistical analysis confirmed that the difference between the highly dependent group (GIR 2-3) and the less dependent groups (GIR 4-5 and GIR 6) was significant.
  • Phoneme Category Impact: Nasal vowels and voiced plosives showed the largest relative differences in alignment scores between aged and non-aged speakers, indicating they are particularly challenging.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that the AGGIR grid is a general measure of autonomy and further work is needed to correlate specific AGGIR variables with ASR performance.
  • While adaptation was effective, the WER for the most dependent elderly speakers (GIR 2-3) remained higher than for less dependent groups.
  • The study focused on French speech; results might differ for other languages.
  • Future Work:
    • Investigate how specific parameters of the AGGIR grid correlate with WER to enable more precise ASR behavior prediction.
  • Develop more sophisticated adaptation techniques that go beyond general acoustic modeling to better address individual variations in aging speech.
  • Explore how these findings can be integrated into AAL systems to improve their reliability and usability for dependent elderly individuals.

Summary updated

Abstract

Ambient Assisted Living aims at providing assistance by allowing people with special needs to perform tasks which they have increasing difficulty with and to provide reassurance through surveillance in order to detect distress and accidental falls. Aged people are among the ones who might benefit from advances in ICT to live as long as possible in their own home. Voice-base smart home is a promising way to provide AAL, but even mature technologies must be evaluated from the perspective of its potential beneficiaries. In this paper, we investigate which characteristics of the ageing voice that challenge a state of the art ASR system. Though in the literature, chronological age is retained as the sole factor predicting decrease in performance, we show that degree of loss of autonomy is even more correlated to ASR performance.
BibTeX
@misc{aman2013analysing,
  author = {Frédéric Aman and Michel Vacher and Solange Rossato and François Portet},
  title = {{Analysing the Performance of Automatic Speech Recognition for Ageing Voice: Does it Correlate with Dependency Level?}},
  year = {2013},
}
Figure