Contribution à l'étude de la variabilité de la voix des personnes âgées en reconnaissance automatique de la parole (Contribution to the study of elderly people's voice variability in automatic speech recognition) [in French]

Frédéric Aman, Michel Vacher, Solange Rossato, François Portet

Abstract: Frédéric Aman, Michel Vacher, Solange Rossato, François Portet. JEP-TALN-RECITAL 2012, Workshop ILADI 2012: Interactions Langagières pour personnes Agées Dans les habitats Intelligents (ILADI 2012: Language Interaction for Elderly in Smart Homes).

2012.

Automatic speech recognition (ASR) systems are increasingly central to Ambient Assisted Living (AAL) technologies, yet they frequently fail the elderly population. This research explores the performance gap between young and elderly speakers and proposes methods to bridge it through acoustic model adaptation.

The Challenge of Aging Voices

Elderly speech is characterized by physiological changes such as vocal cord atrophy and laryngeal calcification, leading to tremors, slower articulation, and imprecise consonant production. Most ASR systems are trained on "standard" voices (healthy adults under 60), creating a significant acoustic mismatch. This study aims to quantify these differences at the phoneme level and evaluate adaptation strategies.

Methodology

The researchers utilized the Sphinx3 ASR engine to perform forced alignment, identifying specific phoneme categories where standard models struggle. To improve performance, they applied Maximum Likelihood Linear Regression (MLLR) to adapt a generic acoustic model (trained on the BREF120 corpus) toward elderly speech characteristics. Furthermore, they investigated the correlation between ASR performance and the AGGIR scale, a French national standard used to measure a person's level of physical and cognitive independence (GIR scores).

Datasets

The study utilized two primary datasets:

  • AD80 (Anondin-Détresse 80): A test corpus of 57 speakers (ages 20–94) reading short home-automation commands and distress calls. This provided the baseline for comparing "non-elderly" vs. "elderly" performance.
  • ERES38 (Entretiens RESidences 38): A specialized corpus containing over 17 hours of recordings from 24 elderly speakers (ages 68–98), including both spontaneous interviews and targeted reading tasks.

Key Insights and Phonetic Analysis

The study revealed that recognition degradation is most severe for unvoiced plosives (like /p/, /t/, /k/) and unvoiced fricatives (like /f/, /s/). As motor control declines with age, these sounds often become "voiced," confusing standard models.

Forced alignment scores comparing non-elderly and elderly voice groups across various phoneme categories.

Interestingly, chronological age was a poor predictor of ASR error rates ($r = -0.053$). Instead, functional autonomy (GIR scores) proved to be the primary driver of voice variability. Speakers with lower autonomy scores exhibited significantly higher error rates than those who were fully independent. This suggests that ASR systems for the elderly should be tuned based on a user's health profile rather than their birth year.

Key Results

Without adaptation, the Word Error Rate (WER) for elderly speakers was a frustrating 27.56%, compared to a crisp 7.33% for the non-elderly group. By applying MLLR adaptation using the ERES38 corpus, the researchers reduced the elderly WER to 11.95%—a 15.6% absolute improvement and a relative error reduction of over 56%.

Word Error Rate (WER) before vs. after acoustic adaptation, showing consistent performance gains.

The adapted model's specificity was also noted; it performed worse on younger voices, confirming that elderly speech occupies a distinct acoustic space that requires specialized modeling.

Conclusion and Future Directions

These findings suggest that ASR systems for seniors should be tuned based on health and autonomy profiles rather than chronological age. Future research directions include integrating non-verbal cues (e.g., hesitations, mouth noises) into acoustic models and exploring ASR performance as a potential early-warning diagnostic tool for detecting declines in a senior's autonomy.

Summary updated

Abstract

Frédéric Aman, Michel Vacher, Solange Rossato, François Portet. JEP-TALN-RECITAL 2012, Workshop ILADI 2012: Interactions Langagières pour personnes Agées Dans les habitats Intelligents (ILADI 2012: Language Interaction for Elderly in Smart Homes). 2012.
BibTeX
@inproceedings{aman2012contribution,
  author = {Frédéric Aman and Michel Vacher and Solange Rossato and François Portet},
  title = {{Contribution à l'étude de la variabilité de la voix des personnes âgées en reconnaissance automatique de la parole (Contribution to the study of elderly people's voice variability in automatic speech recognition) [in French]}},
  booktitle = {ILADI@TALN},
  year = {2012},
  url = {https://openreview.net/forum?id=q2VylpmseL0},
}
Figure