Finnian Kelly

Speaker verification with long-term ageing data

The change experienced by the voice due to ageing must be considered in the development of a long-term speaker verification system. This difficult, largely open, research problem has received little attention to date. For this study, a new Speaker Ageing Database has been collected, containing speech from 18 speakers over a 30-60 year time span. A speaker verification evaluation of this data with a Gaussian Mixture Model - Universal Background Model system reveals that the verification scores of genuine speakers decrease progressively as the time span between training and testing increases, while the imposter scores are less affected. As a consequence, applying a decision threshold fixed at time of enrolment results in a high classification error rate after only a few years. A stacked classifier method of introducing an ageing-dependent decision boundary is applied, significantly improving long-term verification accuracy. Due to score variability at extended time spans however, accurate classification remains a challenging research problem. The ageing-dependent classification approach introduced here represents a first step towards dealing with long-term ageing in speaker verification systems.

Effects of long-term ageing on speaker verification

The changes that occur in the human voice due to ageing have been well documented. The impact of these changes on speaker verification is less clear. In this work, we examine the effect of long-term vocal ageing on a speaker verification system. On a cohort of 13 adult speakers, using a conventional GMM-UBM system, we carry out longitudinal testing of each speaker across a time span of 30-40 years. We uncover a progressive degradation in verification score as the time span between the training and test material increases. The addition of temporal information to the features causes the rate of degradation to increase. No significant difference was found between MFCC and PLP features. Subsequent experiments show that the effect of short-term ageing (<5 years) is not significant compared with normal inter-session variability. Above this time span however, ageing has a detrimental effect on verification. Finally, we show that the age of the speaker at the time of training influences the rate at which the verification scores degrade. Our results suggest that the verification score drop-off accelerates for speakers over the age of 60. The results presented are the first of their kind to quantify the effect of long-term vocal ageing on speaker verification.

Speaker verification in score-ageing-quality classification space

A challenge in automatic speaker verification is to create a system that is robust to the effects of vocal ageing. To observe the ageing effect, a speakers voice must be analysed over a period of time, over which, variation in the quality of the voice samples is likely to be encountered. Thus, in dealing with the ageing problem, the related issue of quality must also be addressed. We present a solution to speaker verification across ageing by using a stacked classifier framework to combine ageing and quality information with the scores of a baseline classifier. In tandem, the Trinity College Dublin Speaker Ageing database of 18 speakers, each covering a 30-60 year time range, is presented. An evaluation of a baseline Gaussian Mixture Model-Universal Background Model (GMM-UBM) system using this database demonstrates a progressive degradation in genuine speaker verification scores as ageing progresses. Consequently, applying a conventional threshold, determined using scores at the time of enrolment, results in poor long-term performance. The influence of quality on verification scores is investigated via a number of quality measures. Alongside established signal-based measures, a new model-based measure, Wnorm, is proposed, and its utility is demonstrated on the CSLU database. Combining ageing information with quality measures and the scores from the GMM-UBM system, a verification decision boundary is created in score-ageing-quality space. The best performance is achieved by using scores and ageing in conjunction with the new Wnorm quality measure, reducing verification error by 45% relative to the baseline. This work represents the first comprehensive analysis of speaker verification on a longitudinal speaker database and successfully addresses the associated variability from ageing and quality arte-facts.

Auditory Features Revisited for Robust Speech Recognition

Auditory based front-ends for speech recognition have been compared before, but this paper focuses on two of the most promising algorithms for noise robustness in automatic speech recognition (ASR). The feature sets are Zero-Crossings with Peak Amplitudes (ZCPA) and the recently introduced Power-Law Nonlinearity and Power-Bias Subtraction (PNCC). Standard Mel-Frequency Cepstral Coefficients (MFCC) are also tested for reference. The performance of all features is reported on the TIMIT database using a HMM-based recogniser. It is found that the PNCC features outperform MFCC in clean conditions and are most robust to noise. ZCPA performance is shown to vary widely with filter bank configuration and frame length. The ZCPA performance is poor in clean conditions but is the least affected by white noise. PNCC is shown to be the most promising new feature set for robust ASR in recent years.

Changes in Speech Chunking in Reading Aloud is a Marker of Mild Cognitive Impairment and Mild-to-Moderate Alzheimer’s Disease

BACKGROUND Speech and Language Impairments, generally attributed to lexico-semantic deficits, have been documented in Mild Cognitive Impairment (MCI) and Alzheimers disease (AD). This study investigates the temporal organisation of speech (reflective of speech production planning) in reading aloud in relation to cognitive impairment, particularly working memory and attention deficits in MCI and AD. The discriminative ability of temporal features extracted from a newly designed read speech task is also evaluated for the detection of MCI and AD. METHOD Sixteen patients with MCI, eighteen patients with mild-to-moderate AD and thirty-six healthy controls (HC) underwent a battery of neuropsychological tests and read a set of sentences varying in cognitive load, probed by manipulating sentence length and syntactic complexity. RESULTS Our results show that Mild-to-Moderate AD is associated with a general slowness of speech, attributed to a higher number of speech chunks, silent pauses and dysfluences, and slower speech and articulation rates. Speech chunking in the context of high cognitive-linguistic demand appears to be an informative marker of MCI, specifically related to early deficits in working memory and attention. In addition, Linear Discriminant Analysis shows the ROC AUCs (Areas Under the Receiver Operating Characteristic Curves) of identifying MCI vs. HC, MCI vs. AD and AD vs. HC using these speech characteristics are 0.75, 0.90 and 0.94 respectively. CONCLUSION The implementation of connected speech-based technologies in clinical and community settings may provide additional information for the early detection of MCI and AD.

PROSODIC CHUNKING IN OUT-LOUD READING: A WINDOW INTO COGNITIVE FUNCTION IN MILD-TO-MODERATE ALZHEIMER’S DISEASE

panel of rat (APPsi) and mouse (Tg4-42) biofluids and tissues. This give further help to understand the devastating neurodegenerative disease, related complex biochemical pathways and pathophysiological processes of AD.Conclusions:UsingNMR-basedmetabolic phenotyping we defined a quantitative readout of transgenic animal models in the form of a biomarker panel. These biomarkers not only contribute to the understanding of this devastating neurodegenerative disease and the related pathophysiological processes on a systemic level, but set the base for a wide range of biomedical applications. Our approach can be easily extended to other tissues, matrices, or disease models and translated across species since metabolic pathways are conserved through evolution, and are essentially similar in rodents and humans.