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

TL;DR

This paper investigates the challenges of using automatic speech recognition (ASR) for elderly individuals, a crucial component for ambient assisted living. By analyzing speech data and adapting acoustic models, the researchers demonstrate significant improvements in ASR performance for older adults, showing that dependency level, not just age, is a key factor in ASR accuracy.

What Problem Does This Paper Address?

As the global population ages, there's a growing need for technologies that support independent living, particularly for those experiencing a decline in autonomy. Ambient Assisted Living (AAL) systems, which often rely on voice interaction for tasks like home automation, are a promising avenue. However, current ASR systems are typically trained on younger adult speech and struggle with the unique characteristics of elderly voices. These characteristics, including vocal tremors, imprecise consonant production, and slower articulation, lead to a significant degradation in ASR performance. This paper aims to bridge this gap by characterizing the specific speech variations in older adults and developing methods to adapt ASR systems to better recognize their speech, thereby enabling more effective AAL solutions.

Methods

The study employs a multi-stage approach:

  • Phoneme Analysis: The researchers first analyze the AD80 corpus to identify which phonemes are most poorly recognized in elderly voices compared to younger adults. This involves using forced alignment to compare speech segments to phonetic models.
  • Acoustic Model Adaptation: To improve ASR performance, they collect a new corpus (ERES38) specifically of elderly speech. They then use Maximum Likelihood Linear Regression (MLLR) to adapt a generic acoustic model (trained on the BREF120 corpus) to this elderly speech data.
  • Correlation Analysis: Beyond age, the paper investigates the correlation between ASR performance and other factors related to elderly individuals' autonomy, using the AGGIR scale (a French assessment tool for dependency).

Datasets

  • AD80 Corpus: This corpus was recorded for the study and contains speech from 57 speakers (20-94 years old), including both younger and older adults. It's specifically designed for home automation and distress call detection. It consists of 6,848 annotated sentences.
  • ERES38 Corpus: This is a newly collected corpus of elderly speech, comprising 17 hours and 44 minutes of recordings from 24 speakers (68-98 years old). It includes both spontaneous speech and text reading.
  • BREF120 Corpus: A large, existing French speech corpus (100 hours) used to train the initial generic acoustic model.

Key Insights and Ideas

  • Age-Related Phoneme Challenges: Elderly voices present specific challenges, with plosives and nasal vowels being particularly difficult for standard ASR systems to recognize accurately.
  • Dependency Over Age: The study reveals that age itself is not the primary driver of ASR performance degradation. Instead, the level of dependency and physical/psychological degradation, as measured by the AGGIR scale, appears to be a stronger indicator of ASR accuracy.
  • Effective Acoustic Adaptation: Even with a relatively small amount of adapted speech data (48 minutes from ERES38), MLLR adaptation significantly reduces the Word Error Rate (WER) for elderly speakers, bringing it close to the performance observed with younger adults using the generic model.
  • Commonalities in Elderly Speech: The success of adapting a model using speakers different from the test set suggests that there are common acoustic characteristics across elderly voices that can be leveraged for adaptation.

Key Results

  • WER Degradation: Without adaptation, the WER for elderly speakers was 27.56%, a significant increase compared to 7.33% for younger adults, representing an absolute difference of 20.23%.
  • Adaptation Improvement: After MLLR adaptation using the ERES38 corpus, the WER for elderly speakers dropped to 11.95%, a relative reduction of 56.65%. This adapted model also showed a WER of 10.39% on younger adults, indicating its specificity to elderly speech.
  • Correlation: The Pearson correlation between WER and age was found to be negligible (-0.053). However, the study suggests a correlation between WER and dependency levels, with lower dependency scores (higher autonomy) generally correlating with lower WER.

Limitations and Future Work

  • Limited Scope of Adaptation: The adaptation was performed globally on the ERES38 corpus. Future work could explore speaker-specific adaptation for even greater personalization.
  • Further Correlation Analysis: While preliminary, the study suggests that specific AGGIR parameters (e.g., motor control of upper limbs, continence) might be highly correlated with WER. Further detailed analysis of these correlations is planned.
  • Non-Verbal Sounds: The paper acknowledges that non-verbal sounds (inspirations, mouth noises) and articulation defects are more frequent in elderly speech and proposes to incorporate these into future acoustic and language models.
  • Predicting WER: A future goal is to develop a system that can predict ASR performance based on an elderly person's dependency characteristics, further facilitating the deployment of AAL technologies.

Summary updated

Abstract

National audience
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