Abstract

The need for cross-lingual information access is more than ever with the easy access to the Internet. This is especially true in vastly multilingual societies like India. Cross-lingual summarization (CLS) aims to create summaries in a target language from a document or document set given in a different, source language. Cross-lingual summarization can have significant impact on enabling cross-lingual information access for millions of people across the globe who do not speak or understand languages that have a large representation on the web by making the most important information available in the target language in the form of summaries. It can also make documents originally published in local languages quickly accessible to a large audience which does not understand those local languages. Working towards a better cross-lingual summarization system, we first create a flexible, web-based tool, referred to as the workbench, for human editing of cross-lingual summaries to rapidly generate publishable summaries in a number of Indian languages for news articles originally published in English. The workbench simultaneously collects detailed logs about the editing process at article, summary and sentence level. Similar to translation post-editing logs, such logs can be used to evaluate the automated cross-lingual summaries in terms of effort needed to make them publishable. We use the workbench to generate two manually edited datasets for different tasks. We observed that quality of automatic translation is a major bottleneck when working on CLS. Translation Quality Estimation (QE) aims to estimate the quality of an automated machine translation (MT) output without any human intervention or reference translation. With the increasing use of MT systems in various cross-lingual applications, the need and applicability of QE systems is increasing. We study existing approaches and propose multiple neural network approaches for sentence-level QE with a focus on MT outputs in Indian languages. For this, we also introduce five new datasets for four language pairs: two for English–Gujarati, and one each for English–Hindi, English–Telugu and English–Bengali, which includes one manually post-edited dataset for English–Gujarati created using the workbench. We compare results obtained using our proposed models with multiple existing state-of-the-art systems including the winning system in the WMT17 shared task on QE and show that our proposed neural model which combines the discriminative power of carefully chosen features with Siamese Convolutional Neural Networks (CNNs) works significantly better for all Indian language datasets. Later, we integrate our efforts on QE with cross-lingual summarization to study its effect on CLS. We extend a popular mono-lingual summarization method to work with CLS, along with a new objective function to take QE scores into account while ranking sentences for summarization. We experimentwith a number of existing methods for CLS with different parameters and settings and show comparative analysis. At the end, we publish an end-to-end CLS software called clstk to make CLS accessible to a larger audience. Besides implementing a number of methods proposed by different CLS researchers over the years, the tool-kit also includes bootstrap code for easy implementation and experimentation with new CLS methods. We hope that this extremely modular tool-kit will help CLS researchers contribute more effectively to the area as well as developers to easily use existing methods in end-user applications.

Abstract

Since the inception of Web 2.0, the virtual data landscape has witnessed unprecedented flow of user generated texts. With applications ranging from product analysis and implementing business strategies to inculcating public opinion in governance processes, Sentiment analysis has become one of the focal points of scientific exploration. Sentiment analysis can be used to decipher subjective opinions expressed over various online platforms using natural language processing methods, statistical and machine learning techniques. With English language having witnessed extensive research in this domain, deciphering the same in vernacular languages remains a prodigious task. While significant strides have been made in Indian languages in the recent years, there has been negligible progress with respect to Sentiment analysis for Odia language, owing to the unavailability of essential tools and resources to identify subjective information in Odia texts. This thesis aims at bridging this gap by building such tools and resources. In order to capture subjective information at sentence-level, a sentiment lexicon becomes a valuable tool. In this thesis, we propose two distinct approaches towards building such a lexicon for resource-poor Indian languages such as Odia. The first methodology involves a translation-based approach which makes use of available resources in English to create a source lexicon. Further using an appropriate machine translation system, the target lexicon is created. A machine translation system between English and the target language being a prerequisite for this approach, the unavailability of the same for Odia language (and a few other resource-poor Indian languages) becomes a methodological drawback. In order to overcome such a disadvantage, we propound a synset-based approach to create the Odia sentiment lexicon. This approach makes use of a linked WordNet structure along with sentiment lexicons available for the used Indian languages. To address the lack of sentiment annotated corpora, we have built two datasets from different domains. Firstly, we have created a corpus of 730 Odia poems annotated with sentimental polarity. An annotation scheme comprising a polarity identification questionnaire clubbed with a taxonomy of emotions has been adopted to label each poem, as positive or negative, at a document level. In order to establish a strong baseline for the corpus, several experiments have been conducted. We employed several machine learning techniques along with word-level features for the initial classification task. Usage of character-level features showed consistent improvementsto the baseline in comparison to word-level features. By creating a word embedding model for Odia language, we have exploited word vector representations as features for classification of Odia poems, the results of which are comparable to that of character-level features. In the domain of news, the thesis also presents a sentiment annotated corpus of 2045 Odia sentences. For sentence-level annotation, proper annotation guidelines have been discussed in order to categorize each sentence as positive, negative or neutral. Baseline results for the corpus have been established with machine learning techniques, using both word and character-level features, for binary and ternary classification. Among these classifiers, Support Vector Machinesand Logistic Regression show comparable results in the classification task. Furthermore, in orderto bear witness to the reliability and utility of the Odia sentiment lexicon, affective words at sentence-level have been identified and used as features for classification. Inclusion of thesefeatures show consistent improvements for both binary and ternary classification in all metrics of evaluation. This testifies the usability of the created Odia sentiment lexicon for sentence-level sentiment classification.

Abstract

Small and medium sized obstacles such as rocks, small boulders, bricks left unattended on the road can pose hazards for autonomous as well as human driving situations. Many times these objects are too small on the road and go unnoticed on depth and point cloud maps obtained from state of the art range sensors such as 3D LIDAR. In the first part of thesis, a novel algorithm is proposed that fuses both appearance and 3D cues such as image gradients, curvature potentials and depth variance into a Markov Random Field (MRF) formulation that segments the scene into obstacle and non obstacle regions. Appearance and depth data obtained from a ZED stereo pair mounted on a Husky robot is used for this purpose. While identifying true positive obstacles such as rocks, large stones accurately our algorithm is simultaneously robust to false positive sources such as appearance changes on the road, papers and road markings. High accuracy detection in challenging scenes such as when the foreground obstacle blends with the background road scene vindicates the efficacy of the proposed formulation. In the second part of thesis, a novel neural network architecture called MergeNet for discovering small obstacles for on-road scenes in the context of autonomous driving is proposed. The basis of the architecture rests on the central consideration of training with less amount of data since the physical setup and the annotation process for small obstacles is hard to scale. For making effective use of the limited data, we propose a multi-stage training procedure involving weight-sharing, separate learning of low and high level features from the RGBD input and a refining stage which learns to fuse the obtained complementary features. The model is trained and evaluated on the Lost and Found dataset and is able to achieve state-of-art results with just 135 images in comparison to the 1000 images used by the previous benchmark. Additionally, we also compare our results with recent methods trained on 6000 images and show that our method achieves comparable performance with only 1000 training samples.

Abstract

Motion planning is an important primitive for a robotic or mechanical system with closed kinematic chains that lets the robot find the shortest or otherwise optimal path in order to reach its goal position. Sampling based methods provide an efficient solution for motion planning in basic robot systems. However, complex robotic systems such as humanoids have a large number of degrees- of-freedom (DoF) which makes motion planning extremely dicult. Further, these systems also have constraints such as self-collision avoidance and active balancing. Thus, incorporating these constraints in existing motion planners is complex and computationally expensive. Therefore, real time calculation of inverse kinematics (IK) with dynamically stable configuration is of high necessity in humanoid robots as they are highly susceptible to lose balance. In the first part of this thesis, a methodology to generate joint-space trajectories of stable configurations for motion planning using Deep Reinforcement Learning (RL) is proposed. The methodology is based on Deep Deterministic Policy Gradient (DDPG), where the robot autonomously learns the optimal behavior through a series of trial-and-error interaction with the environment. The proposed strategy was evaluated for various motion planning tasks on the highly articulated upper body of a humanoid model with 27 degree-of-freedom (DoF). Most reinforcement learning algorithms are inefficient for learning multiple tasks in complex robotic systems, where actions between different tasks are not fully separable and share a set of actions. For cases where task-specific actions are fully separable, each task policy can be learnt independently using different policy networks. However when the task-specific actions are not fully separable, the policies cannot be learnt independently. In such environments, a compound policy may be learnt with shared neural network parameters which performs multiple tasks concurrently. However such compound policy may get biased towards a task or the gradients from dierent tasks negate each other, making the learning unstable and sometimes less data efficient. The main contribution of this work is a novel framework for simultaneous learning of multiple tasks sharing a set of common actions in continuous action spaces, which we call as DiGrad (Differential Policy Gradient). The proposed framework is based on differential policy gradients (DPG) and can accommodate multi-task learning in a single actor-critic network. We also propose a simple heuristic in the differential policy gradient update in case of partially separable actions to further improve learning. The experimental results show that the framework supports ecient multi-task learning in complex robotic systems, outperforming related methods in continuous action spaces such as DDPG. With the advent of artificial intelligence and machine learning, humanoid robots are made to learn a variety of skills which humans possess. One of fundamental skills which humans use in day-to-day activities is performing tasks with coordination between both the hands. In case of humanoids, learning such skills require optimal motion planning which includes avoiding collisions with the surroundings. In the nal part of this thesis, a methodology based on DiGrad to learn coordinated tasks in cluttered environments is proposed. Further, we propose an algorithm to smooth the joint space trajectories obtained by the proposed framework in order to reduce the noise instilled during training. The proposed framework was tested on a 27 DoF humanoid with articulated torso for performing coordinated object-reaching task with both the hands in four different environments with varying levels of difficulty.

Abstract

The rapid advancements in the field of semiconductor technologies has paved the way for high speed embedded systems that require high speed of operation, lower power dissipation and robust to electromagnetic noises and interference. The synchronous digital system have been a successful alternative due to simplicity of design but the advancements in the embedded system has imposed major challenges in the simultaneous reduction of delay and power consumption. The reduction in clock period in synchronous design enables higher operating frequency that enhances the clock synchronization complexity and dynamic power consumption leading to the surge in the demand of asynchronous architecture. The clock routing and power consumption issue in synchronous design have been eradicated with the concept of (globally asynchronous locally synchronous) GALS architecture in the year 1975’s. There was no global clock in the design and the sub-modules in the design interact with each other using asynchronous handshaking protocols. The highest operating frequency in GALS architecture was still constrained and limited with respect to the slowest sub-module clock frequency. The fully asynchronous architecture eradicates this limitation since the data transfer is irrespective of clock transitions and depends upon the availability of valid data in order to transfer from source to receiver end. In this thesis, we have proposed the design of modified 4-phase handshaking protocol on the lines of existing standard 4-phase protocol developed earlier for data communication in asynchronous circuits. The proposed protocol has achieved better performance in terms of speed by taking 1-cycle lesser than the existing 4-cycle implementation of standard 4-phase with minimal area overhead. The simulation results obtained for the design of asynchronous ALU based on modified 4-phase handshaking protocol performs better than the existing synchronous and asynchronous standard 4-phase design by executing higher number of instructions in a given time. The time taken in simulation experiment was computed by taking fixed data set of random instructions including ADD, SUB, OR, XOR etc that showed 1.4x increase in the performance in comparison with the asynchronous ALU implemented using standard 4-phase protocol design. The simulation result for the design of asynchronous FSM based on modified 4-phase handshaking protocol outperforms the number of bit sequences detected in a non-overlapping fashion(speed) by existing synchronous and asynchronous standard 4-phase design. The results indicate an increase in speed of 1.65x as compared to the synchronous design and 1.4x in comparison with the design based on asynchronous standard 4-phase protocol.

Abstract

Wireless devices are typically powered by batteries to ensure their portability. The finite energy storage capacity of the batteries limits operational period of the network. Recharging and replacing batteries of devices is time consuming or expensive, particularly when the sensors are deployed in hostile or hard-to-access environments. Energy harvesting (EH), a technique to collect energy from the ambient energy sources including radio frequency (RF) signals, has received considerable attention as a viable solution to this problem. Simultaneous wireless information and energy transfer (SWIET) explores a dual use of RF signals to transfer information jointly with energy using the same waveform. This can guarantee a reliable supply of energy unlike the case of other ambient energy sources. In this work, we consider SWIET in a cooperative wireless network involving two transceiver nodes whose communication is assisted by energy-constrained two-way amplify-and-forward (AF) relay nodes. The transceiver nodes can simultaneously transmit information and energy, and the relay nodes harvest the energy from the received signal and use the harvested energy to amplify and retransmit the received signal to the transceiver nodes. We consider two practical receiver architectures, namely, time switching (TS) and power splitting (PS). In TS, the receiver harvests energy for a fraction of transmission block and then decodes information for the remaining fraction. In PS, a fraction of the received signal is used for energy harvesting and the rest for information decoding. Based on the TS and PS receiver architectures, we propose the TSR and PSR protocols to facilitate simultaneous energy harvesting and information processing at the relay. For both protocols, we address the problem of resource allocation along with energy harvesting and optimal relay selection. The performance of the proposed PSR and TSR protocols for different parameters is illustrated by numerical simulations. We extend the relaying protocols discussed above to a cognitive radio (CR) scenario. Cognitive radio networks (CRNs) achieve better spectral efficiency by allowing the licensed users to share their spectrum with unlicensed users. The interference produced by the unlicensed users is kept below an accepted threshold. We consider a CRN with a licensed primary user (PU), two secondary transceivers, and a set of secondary two-way amplify and forward (AF) relays. The secondary user (SU) nodes share the spectrum with the PU node, and each node is assumed to be equipped with a single antenna. SU transceivers communicate with each other only via the relay optimally selected form the set of available relays. The transmit signals from the SU transceivers are designed so as to carry both information and energy. The relay harvests energy from these signals and uses this energy to process the received signals for further transmission.

Abstract

This thesis aims to investigate different aspects of timing in Sequence Learning and Adaptation to certain emotional responses. Sequencing and processing of emotions are fundamental to human behaviour and hence are important fields to explore. To investigate the influence of temporal factors in Sequence learning, we conducted a behavioural study using the paradigm of Serial Reaction Time (SRT) task and introduced delay in presentation of successive stimuli, referred to as Response-to-Stimulus Interval (RSI). By varying the RSI between suc- cessive elements of a sequence over a few temporal windows, we attempted to explore what effects this systematic varying of the RSI would have on the sequence learning process. We have also con- ducted post-experimental debriefing sessions to see if such learning is implicit or explicit. The SRT task requires the participants to simply respond to a stimulus that appears on the screen by clicking a key corresponding to the location at which it appeared. Reaction times were recorded for all the stimuli and reduction in this would imply learning. Statistical analysis of the reaction time data showed that learning did not get impaired in the various temporal groups despite delays experienced due to vari- able RSI. Post-experimental results showed that the knowledge acquired in smaller RSIs was implicit while for larger RSIs it was explicit. The results are in general agreement with the existing literature with fixed RSIs and hence we can conclude from the empirical study that varying RSIs do not impair sequence learning and that the sequential knowledge might have become explicit with larger RSIs. Ad- ditionally, we proposed a biologically realistic computational model in the form of a simple recurrent neural network (SRN) to obtain a functional account of the empirical findings. The proposed model has spatial representation of time for incorporating RSI in the neural network. The model illustrates how explicit learning could emerge due to a longer temporal window between stimuli that could potentially give insights into the mechanisms of sequence learning in variable RSI conditions. In the second part of the thesis, we investigated the duration aspect of timing. Specifically, how different kinds of emotional reactions last for different duration. For this, we conducted an empirical study with a sad and self-relevant stimulus and compared the results to a sad but not-so-self-relevant stimulus. We found that the emotional reaction in the current study lasted longer because it is perhaps difficult for human beings to adapt to an emotional response when they can relate to the stimulus that elicited the emotion. The emotional intensity and its duration are thus determined by two factors, the self-relevance of the affective stimulus and how well or poorly understood the stimulus is to the person experiencing it. The greater the self-relevance and poorer the understanding, the stronger the emotion and longer will be its duration. We also describe a preliminary study conducted to collect Physiological data as an attempt to strengthen this observation. The thesis ends with a summary of the work and a general discussion of the findings. Limitations and future work have also been discussed.

Abstract

With the evolution of wireless systems, more communication standards are being proposed while maintaining backward compatibility, therefore, there is an ever growing need for wideband multistandard receivers similar to software defined radio (SDR) receivers. For many years now, SDR or digitally reconfigurable radio research has been quite challenging with only a few reported practical prototypes with limited success. The basic demand of the SDR has been the need for a giant analog-to-digital converter (ADC) and a powerful enough digital signal processing (DSP) so that it serves as a universal radio platform receiving almost all radio standards and services. These demands are difficult to meet even in modern CMOS technologies leading to the necessity of highly flexible architectures compared to traditional receivers. To achieve this objective, digital intensive receiver architectures with passive structures are proposed. Radio frequency (RF) sampling receivers using time-interleaved (TI), RF sampling and discrete-time mixing also provide additional flexibility by employing discrete-time signal processing in analog domain. Owing to the oversampling ratio of up to 8 times in these receivers, high-performance complex frequency synthesizers are needed with stringent phase noise requirement. Moreover, the clocking schemes employed in TI RF sampling architectures result in additional timing jitter and mismatch offsets. These challenges are alleviated in a class of sampling receivers, called subsampling receivers, where the employed bandpass sampling principle results in less than the Nyquist sampling frequency. This leads to the less-complex and less power hungry frequency synthesizers. In this thesis, an intermediate solution to SDR receiver implementation called “Mini-SDR” architecture is proposed by exploiting the bandpass sampling concept. Based on the proposed Mini-SDR architecture, three CMOS radio frontends are proposed for standalone electronic attack (EA) transceiver system, dynamically reconfigurable multistandard subsampling (DRMS) radio receiver and a dual-band subsampling receiver for IEEE WLAN 802.11ac standard. The first architecture is a single chip integrated transceiver for a standalone EA system based on digital radio frequency memory (DRFM) repeater without the need for a separate instantaneous frequency measurement (IFM) receiver. The proposed transceiver architecture employs sub-Nyquist discrete-time frontend to estimate the carrier frequency and bandwidth of the incoming pulse radar signals. The system is verified for linear frequency modulated, constant carrier frequency and phase coded pulse radar signals. These signals are classified on the basis of Wigner-Ville distribution. The performance of the proposed system is validated from 1 GHz to 20 GHz with instantaneous bandwidth of 300 MHz, dynamic range from −60 dBm to −20 dBm and system gain from 30 dB to 60 dB. The minimum delay achieved between reception and retransmission is 13 pulse repetition intervals (PRIs). To the best of author’s knowledge, this work reports the first sub-Nyquist bandpass sampling based single chip architectural solution for a standalone EA system where RF frontend and digital controller are integrated into the system. The second architecture is the DRMS radio receiver. The proposed receiver has a unique capability to detect the carrier frequency of the incoming signal, estimate it’s bandwidth and standard. The RF frontend is modelled in Verilog-AMS behavioural model and the digital signal processing is implemented in Simulink-Matlab. Also, the system level optimizations of the overall receiver performance are presented with the effect of receiver impairments. The complete receiver architecture has been verified to detect and process three different bands belonging to three different standards (GSM, UMTS and WLAN) with the carrier frequency ranging from 0.9 GHz to 2.5 GHz with a maximum signal bandwidth of 22 MHz and the input dynamic range from −109 dBm to −20 dBm. The last implementation is a first dual-band subsampling receiver with subsampling frequency optimization to meet ultimate receiver error vector magnitude (EVM) of −40 dB over wide input power range of 19 dB. Systematic system level optimization of the receiver chain with major impairments such as sampling frequency, synthesizer phase noise, IQ mismatch and unit capacitor in the SC filter is also proposed. Sampling frequency optimization proposed in this work has multi-fold advantages: a) the noise folding effect is reduced leading to smaller noise figures. b) less-complex and less-power hungry clocking scheme. c) quadrature phase readily available between consecutive samples is utilized to separate samples into in-phase (I) and quadrature-phase (Q) paths. d) the ultimate EVM performance of the receiver is significantly better than the subsampling receivers, where the sampling frequency is not optimized. The dual-band subs

Abstract

Decentralized MDPs (Dec-MDPs) provide a rigorous framework for collaborative multi-agent sequential decision making under uncertainty and partial observability. However, their high computational complexity limits their practical impact. To overcome this computational complexity, we focus on a special class of DEC-MDPs consisting of independent agents which collaborate with each other through a global joint-reward function that depend upon their entire histories of states and actions to accomplish joint tasks. We make the following contributions to address the issue of scalability for this class of problems. 1) Dec-NLP: A nonlinear programming (NLP) formulation for such event-based planning model which scales well for small problems. However, we observed empirically that with the increasing number of agents, NLP solvers were unable to scale, slow and often ran out of memory; 2) Dec-EM: A probabilistic inference based approach that scales much better than NLP solvers for a large number of agents. However, the scalability still suffers when the state-space of each agent is exponential, which is often the case for several patrolling and coverage problems; 3) Dec-MARL: A policy gradient based multi-agent Reinforcement Learning (RL) approach that scales well even for exponential state-spaces. (4) Dec-ESR : A new actor-critic based Reinforcement Learning (RL) approach for event based Dec- MDPs that produces better solution quality than Dec-MARL using successor features (SF) which is a value function representation that decouples the dynamics of the environment from the rewards. We then present how Dec-ESR generalizes learning for event based Dec-MDPs using SF within an end-to-end deep RL framework; (5) We show that the proposed method using SF allows useful transfer of information on related but different tasks, hence bootstraps the learning for new tasks and makes convergence much faster on new tasks. Our inference and RL-based advances enable us to solve a large real-world multi-agent coverage problem modeling schedule coordination of agents in a real urban subway network where other approaches fail to scale.

Abstract

Speech often consists of expressive and nonverbal sounds in addition to the message to be conveyed. The objective of this thesis is to analyze expressive voices to determine the features needed for representing, describing and discriminating different types of expressions. Here expression refers to the information in the signal produced by the speech production mechanism, that characterizes different phonations, emotions, singing (normal and artistic) and nonverbal sounds. Most of these expressive features are as a result of the variations in the excitation component of speech production processes, although it is difficult to isolate the excitation component from the dynamic vocal tract system component. Also, the expressiveness is reflected in the dynamics of the excitation component. Few methods for extracting the excitation source features rely on inverse filtering methods, which are known to have limitation for high pitched voices such as emotional speech and singing voice. In this work, we propose a representation of the excitation source in terms of impulse-like sequence to capture variations in the source. Significance of the excitation source is studied for three types of expressive voices, namely, phonation types (in speech and singing), emotional speech and artistic (Noh) singing voice. The primary mode of excitation is due to vibration of the vocal folds at the glottis. Even though the excitation information is present throughout the glottal cycle, it is considered to be significant only when there is large change in a short-time interval, i.e., when it is impulse-like. The impulse-like characteristic is usually exhibited around the instant of glottal closure within each glottal cycle. Apart from the primary impulse-like excitation at the instant of glottal closure, other major and minor excitation impulses may also occur at glottal opening and in creaky voice. Other excitation impulses become prominent in aperiodic voices, like artistic (Noh) singing voice. An approach for extracting the excitation source information in terms of impulse-like sequence is proposed using single frequency filtering (SFF) method. To exploit the impulse-like discontinuities from the signal, impulse properties in magnitude and phase component are explored. An impulse-like characteristic is derived using the spectral variance computed at each instant of time, and the instantaneous frequency (IF) of the filtered signal at each frequency. The slope of the variance plot shows discontinuities at the locations of the impulse-like excitation. The sum of the IFs computed for each SFF output signal shows discontinuities at the locations of the impulse-like excitations. Extraction of the impulse-like sequence is assessed using synthetic and natural speech signals. The SFF-based method also brings out the minor impulse-like excitations within each glottal cycle. Determination of two important features of excitation, namely, glottal closure instant (GCI) and glottal open region (GOR) are derived by exploiting the changes in the spectral characteristics of the SFF spectra and HNGD spectra of the zero time windowing (ZTW) method. A spectral flatness parameter derived from the SFF spectra highlights the impulse-like characteristics at the GCI. The spectral flatness parameter derived from the HNGD spectra highlights the GOR. A new approach is proposed for detection of GCIs from emotional speech. The approach uses the recently proposed modified zero frequency filtering (mZFF) and the impulse-like sequence representation obtained using the SFF method. Using the features derived from the three signal processing methods namely, ZFF, ZTW and SFF, expressive voices such as phonation types, emotional speech and artistic (Noh) singing voice, are analyzed. Features such as locations of GCIs, strength of the glottal closure, different phases (closed and open phase) of glottal vibration and their spectral correlates, are used for analysis and detection of different phonation types in speech and singing. Cepstral coefficients derived from the ZTW (ZTWCC) and SFF (SFFCC) methods are used for discriminating different phonation types. Excitation features derived around the GCIs using ZFF method and linear prediction (LP) analysis are used to study the role of excitation information in emotional speech. One of the main features of the artistic (Noh) singing voice is the aperiodicity in the vocal fold vibration. The aperiodicity characteristics are studied in detail using the spectral characteristics of the impulse-like sequence, which shows harmonics and subharmonics of the fundamental frequency corresponding to pitch. The significance of the impulse-like sequence representation is also studied through the saliency measure. The saliency measure captures the pitch perception information in the aperiodicity regions. The key contributions of the thesis are: • Analysis of expressive voices using the recently proposed signal processing meth