Electromagnetic articulography (EMA) is a crucial tool in speech production research, enabling the measurement of articulator movements (e.g., tongue, lips, jaw) in real-time. It helps analyze speech patterns in both healthy individuals and those with speech disorders, uncovering the relationships between articulatory movements, acoustic properties, and linguistic structures. EMA data also supports the development of acoustic-to-articulatory inversion (AAI) systems, which convert acoustic signals into articulatory movements.

Sleep is controlled by the body's clock, the Circadian Rhythm. It is this continuous rhythm that alternates wakefulness and drowsiness in the day and night times, respectively. This rhythm can be altered with change in the timings of the activities performed by the person. Sleep has been shown as an important factor that determines the effectiveness of the work performed through the day. To deprive the body of healthy sleep means to deprive the body of its rest quotient. Healthy sleep helps rejuvenate the body, repair tissues, grow muscles, synthesize hormones, retain information better and perform well on memory related tasks. The process of sleeping consists of different sub-levels of wakefulness and sleep - REM and NREM. The study of sleep is usually done using Polysomnograph (PSG) signals acquired using bio-signal electrodes patched to a sleeping person. Through analysis of PSG signals, it is possible to identify sleep disorders like Insomnia, Sleep Apnea, Narcolepsy, etc.
Analysis of long duration PSG graphs is usually carried out by clinical experts who manually study the graph to label sleep stages, arousals and apnea. This project concerns the automation of the same task of characterizing signal patterns to classify sleep stages and detect abnormalities in the sleep cycle, using deep learning techniques.

Patients with Head and Neck Cancer can experience dysphagia or more commonly known as swallowing disorders. Dysphagia can potentially lead to increased-risk medical conditions like pulmonary aspiration (both silent and overt), choking, fatigue and malnutrition. Such cases when unidentified can be fatal. The diagnosis of such disorders involves screening tests, and clinical assessments like fibreoptic endoscopy and videofluoroscopy. These techniques are invasive to the subject and can sometimes be harmful due to exposure to X-rays. Hence, in this project, the assessment of the physiology of swallowing is done using cervical auscultation (CA), where the swallowing action is captured by the sounds that are produced when the food bolus is swallowed. In CA, the sounds are picked up using an external stethoscope or microphone. The aim is to develop automatic methods to study the characteristics of swallow in both healthy and dysphagic subjects and also to detect the severity of dysphagia in Head and Neck Cancer patients, by leveraging signal processing and machine learning methods.
Initial work in this project involved feature learning for a volume dependent analysis and classification of water swallowed. Results indicated that, across different volumes, the acoustic features selected using automatic feature selection methods were more robust to volume changes than the baseline features that pertain to basic temporal and spectral parameters.

While speaking at different rates, articulators (like tongue, lips) tend to move differently and the enunciations are also of different durations. In the past, affine transformation and DNN have been used to transform articulatory movements from neutral to fast(N2F) and neutral to slow(N2S) speaking rates. In this work, we improve over the existing transformation techniques by modeling rate specific durations and their transformation using AstNet, an encoder-decoder framework with attention. In the current work, we propose an encoder-decoder architecture using LSTMs which generates smoother predicted articulatory trajectories. For modeling duration variations across speaking rates, we deploy attention network, which eliminates the need to align trajectories in different rates using DTW. We perform a phoneme specific duration analysis to examine how well duration is transformed using the proposed AstNet. As the range of articulatory motions is correlated with speaking rate, we also analyze amplitude of the transformed articulatory movements at different rates compared to their original counterparts, to examine how well the proposed AstNet predicts the extent of articulatory movements in N2F and N2S. We observe that AstNet could model both duration and extent of articulatory movements better than the existing transformation techniques resulting in more accurate transformed articulatory trajectories.

During advanced stages of cancers of the larynx and hypopharynx, a surgical procedure called laryngectomy is performed, in which cancerous regions, including the larynx or the voice box, are removed. In laryngectomees, in the absence of vocal folds, it is the vibration of the esophagus that gives rise to a low-frequency pitch during speech production, this speech is referred to as Tracheoesophageal speech. The tracheoesophageal (TE) speech is known to be hoarse,breathy and rough, which makes it difficult to understand. This work is about converting TE-speech to natural-sounding speech.

As the heart contracts and relaxes, the blood volume in the facial skin capillaries increases and decreases respectively. This causes a rhythmic variation in the pixel intensity values in the consecutive frames of the recorded facial video. Using this information, signal processing approaches as well as deep learning approaches were applied on the datasets recorded in IISc, to achieve a best mean absolute error of 2 BPM in a one minute video. The results were observed to be strongly varying with the subject and the light conditions.

This study analyzes how the human voice matures with age.

This study analyzes the effect of consonant context and speaking rate on vowel space and coarticulation in Toda vowel-consonant-vowel (VCV) sequences. All vowels /a/,/e/, /i/, /o/, /u/, and two intervocalic consonants, /p/ (labial) and /t/ (alveolar), are considered to form asymmetrical VCV sequences in slow and very fast speaking rates. Acoustic analysis using first and second formants (F1 and F2) shows a significant change in vowel space across speaking rates in a consonant specific manner. Quantification of range and extent of coarticulation using F2 are presented to carry out acoustic analysis of both anticipatory and carryover coarticulation. Significant V-V carryover coarticulation is observed in few VCV sequences with labial consonant. However, significant effect of consonant context is found for both anticipatory and carryover coarticulation in most of the VCV sequences. Increase in speaking rate is found to significantly drop both anticipatory and carryover co-articulation range in the context of alveolar consonants. Results from these acoustic analyses indicate that there are differences in the nature in which rate and consonant context affect the coarticulatory organization.

In speech production, vocal folds play a vital role in modulating airflow from lung through its quasiperiodic vibration. In between the vocal folds, a narrow opening is present which allows air to pass through the trachea called as glottis. The changes in shape or muscular properties of vocal folds lead to changes in the glottis opening. This causes variation in the voice like hoarseness and dysphonia. Many of these lead to an incomplete closure of the glottis during production of voice, this is termed as glottic chink. The glottic chink is typically visualized utilizing endoscopic or stroboscopic video by speech language pathologists (SLPs). This project aims at automatically localizing and segmenting the glottis in a stroboscopic video.

Current clinical methods of diagnosis for asthma are evidently tedious, expensive and time-consuming. The motivation behind Asquire comes from developing a diagnosis method that is easy, yet effective and fast, using vocal sounds powered by Machine Learning (ML) and signal processing techniques.

A Text-to-Speech (TTS) corpus is a collection of recorded speech and corresponding text transcriptions used to develop and train TTS systems. These corpora are essential for creating high-quality TTS engines capable of generating natural-sounding speech from text input.
Please contribute your voice and help us improve our analyses.