Dementia affects the patient’s memory and leads to language impairment. Research has demonstrated that speech and language deterioration is often a clear indication of dementia and plays a crucial role in the recognition process. Even though earlier studies have used speech features to recognize subjects suffering from dementia, they are often used along with other linguistic features obtained from transcriptions. This study explores significant standalone speech features to recognize dementia. The primary contribution of this work is to identify a compact set of speech features that aid in the dementia recognition process. The secondary contribution is to leverage machine learning (ML) and deep learning (DL) models for the recognition task. Speech samples from the Pitt corpus in Dementia Bank are utilized for the present study. The critical speech feature set of prosodic, voice quality and cepstral features has been proposed for the task. The experimental results demonstrate the superiority of machine learning (87.6 percent) over deep learning (85 percent) models for recognizing Dementia using the compact speech feature combination, along with lower time and memory consumption. The results obtained using the proposed approach are promising compared with the existing works on dementia recognition using speech.

We propose a hybrid network-based learning framework for speaker-adaptive vocal emotion conversion, tested on three different datasets (languages), namely, EmoDB (German), IITKGP (Telugu), and SAVEE (English). The optimized learning model introduced is unique because of its ability to synthesize emotional speech with an acceptable perceptive quality while preserving speaker characteristics. The multilingual model is extremely beneficial in scenarios wherein emotional training data from a specific target speaker are sparsely available. The proposed model uses speaker-normalized mel-generalized cepstral coefficients for spectral training with data adaptation using the seed data from the target speaker. The fundamental frequency (F0) is transformed using a wavelet synchrosqueezed transform prior to mapping to obtain a sharpened time–frequency representation. Moreover, a feedforward artificial neural network, together with particle swarm optimization, was used for F0 training. Additionally, static-intensity modification was also performed for each test utterance. Using the framework, we were able to capture the spectral and pitch contour variabilities of emotional expression better than with other state-of-the-art methods used in this study. Considering the overall performance scores across datasets, an average melcepstral distortion (MCD) of 4.98 and root mean square error (RMSE-F0) of 10.67 were obtained in objective evaluations, and an average comparative mean opinion score (CMOS) of 3.57 and speaker similarity score of 3.70 were obtained for the proposed framework. Particularly, the best MCD of 4.09 (EmoDB-happiness) and RMSEF0 of 9.00 (EmoDB-anger) were obtained, along with the maximum CMOS of 3.7 and speaker similarity of 4.6, thereby highlighting the effectiveness of the hybrid network model.

The paper describes the creation, analysis and validation of a multilingual Dementia Speech dataset for Indic languages. Three popular Indian languages viz. Telugu, Tamil and Hindi are considered for the pilot study. Dementia and associated Alzheimers disease affect a large section of Asian population. Though there are promising studies in dementia detection focussed on Western ethnicity, the absence of a clinical dementia dataset for Indian languages forms the primary motivation for this study. This pilot study aims to overcome the challenges associated with data collection and validation in a clinical setting and deal with situations wherein clinical data is not readily available. The Indic dementia dataset is an enacted non-clinical dataset created from the manual translations of the benchmark clinical English DementiaBank dataset. The dataset created is validated using features extracted from the benchmark. The feature evaluation revealed a similarity of 92.6% for silences, 92% for mean pitch (Hz), 84.7% for jitter and 90.3% for shimmer. Subjective evaluation was also conducted based on clarity and similarity of utterances with DementiaBank data. An average MOS of 3.9 for clarity of speech and 3.76 for similarity with respect to DementiaBank was obtained across all three languages. A baseline classification using state-of-art deep network architecture gave a maximum of 78% accuracy in dementia detection using the Indic dementia dataset. The pilot experimentation in this work gives promising insights into the development of a multilingual dataset for analysis of clinical speech patterns in early dementia in the Indian population.

Expressive speech can be synthesized using acoustic feature modelling by mapping the spectral and fundamental frequency parameters between neutral speech and target emotions based on context. Speaker and text-independent emotion conversion are challenging modelling problems in this paradigm. In this paper, spectral mapping using an i-vector based framework of fixed dimensions is proposed for speaker-independent emotion conversion, considering the entire problem in the utterance domain, rather than existing approaches using frame-level processing. The high dimensionality of i-vectors and reduced utterances for i-vector training necessitate the use of Probabilistic Linear Discriminant Analysis (PLDA) to derive the emotion dependent latent vector. The i-vector set-up does not require parallel data or alignment procedures at any stage of training. F0 training is conducted on a multilayer feed-forward neural network using limited aligned seed parallel data. The framework is tested on three different languages (datasets) viz. German (EmoDB), Telugu (IITKGP) and English (SAVEE). The proposed approach delivered superior performance compared to the baseline under both clean and noisy data conditions considered for analysis. Under clean data conditions, the proposed model was found to perform better than the baseline with a Mel Cepstral Distortion as low as 3.8 (fear), an F0-RMSE of 26.31 (happiness) and a Perceptual Evaluation of Speech Quality (PESQ) of 3.64 (anger) across datasets. Subjective testing provided a maximum CMOS of 4.10 (anger), 4.44 (fear) and 3.43 (happiness).