This thesis presents a system for the interpretation of natural speech which serves as input module for a spoken dialog system. It carries out the task of extracting application-specific pieces of information from the user utterance in order to pass them to the control module of the dialog system. By following the approach of integrating speech recognition and speech interpretation, the system is able to determine the spoken word sequence together with the hierarchical utterance structure that is necessary for the extraction of information directly from the recorded speech signal. The efficient implementation of the underlying decoder is based on the powerful tool of weighted finite state transducers (WFSTs). This tool allows to compile all involved knowledge sources into an optimized network representation of the search space which is constructed dynamically during the ongoing decoding process. In addition to the ...

One desideratum in designing cognitive robots is autonomous learning of communication skills, just like humans. The primary step towards this goal is vocabulary acquisition. Being different from the training procedures of the state-of-the-art automatic speech recognition (ASR) systems, vocabulary acquisition cannot rely on prior knowledge of language in the same way. Like what infants do, the acquisition process should be data-driven with multi-level abstraction and coupled with multi-modal inputs. To avoid lengthy training efforts in a word-by-word interactive learning process, a clever learning agent should be able to acquire vocabularies from continuous speech automatically. The work presented in this thesis is entitled \emph{Constrained Non-negative Matrix Factorization for Vocabulary Acquisition from Continuous Speech}. Enlightened by the extensively studied techniques in ASR, we design computational models to discover and represent vocabularies from continuous speech with little prior knowledge of the language to ...

Turkish, being an agglutinative language with rich morphology, presents challenges for Large Vocabulary Continuous Speech Recognition (LVCSR) systems. First, the agglutinative nature of Turkish leads to a high number of Out-of Vocabulary (OOV) words which in turn lower Automatic Speech Recognition (ASR) accuracy. Second, Turkish has a relatively free word order that leads to non-robust language model estimates. These challenges have been mostly handled by using meaningful segmentations of words, called sub-lexical units, in language modeling. However, a shortcoming of sub-lexical units is over-generation which needs to be dealt with for higher accuracies. This dissertation aims to address the challenges of Turkish in LVCSR. Grammatical and statistical sub-lexical units for language modeling are investigated and they yield substantial improvements over the word language models. Our novel approach inspired by dynamic vocabulary adaptation mostly recovers the errors caused by over-generation and ...

Above 20% European citizens speak in Russian therefore the task of automatic recognition of Russian continuous speech has a key significance. The main problems of ASR are connected with the complex mechanism of Russian word-formation. Totally there exist above 3 million diverse valid word-forms that is very large vocabulary ASR task. The thesis presents the novel HMM-based ASR model of Russian that has morphemic levels of speech and language representation. The model includes the developed methods for decomposition of the word vocabulary into morphemes and acoustical and statistical language modelling at the training stage and the method for word synthesis at the last stage of speech decoding. The presented results of application of the ASR model for voice access to the Yellow Pages directory have shown the essential improvement (above 75%) of the real-time factor saving acceptable word recognition rate ...

This thesis investigates the application of language models based on semantic similarity to Automatic Speech Recognition for meetings. We consider data-driven Latent Semantic Analysis based and knowledge-driven WordNet-based models. Latent Semantic Analysis based models are trained for several background domains and it is shown that all background models reduce perplexity compared to the n-gram baseline models, and some background models also significantly improve speech recognition for meetings. A new method for interpolating multiple models is introduced and the relation to cache-based models is investigated. The semantics of the models is investigated through a synonymity task. WordNet-based models are defined for different word-word similarities that use information encoded in the WordNet graph and corpus information. It is shown that these models can significantly improve over baseline random models on the task of word prediction, and that the chosen part-of-speech context is ...

Con dence measures for the results of speech/speaker recognition make the systems more useful in the real time applications. Con dence measures provide a test statistic for accepting or rejecting the recognition hypothesis of the speech/speaker recognition system. Speech/speaker recognition systems are usually based on statistical modeling techniques. In this thesis we de ned con dence measures for statistical modeling techniques used in speech/speaker recognition systems. For speech recognition we tested available con dence measures and the newly de ned acoustic prior information based con dence measure in two di erent conditions which cause errors: the out-of-vocabulary words and presence of additive noise. We showed that the newly de ned con dence measure performs better in both tests. Review of speech recognition and speaker recognition techniques and some related statistical methods is given through the thesis. We de ned also ...

Speech is at the core of human communication. Speaking and listing comes so natural to us that we do not have to think about it at all. The underlying cognitive processes are very rapid and almost completely subconscious. It is hard, if not impossible not to understand speech. For computers on the other hand, recognising speech is a daunting task. It has to deal with a large number of different voices "influenced, among other things, by emotion, moods and fatigue" the acoustic properties of different environments, dialects, a huge vocabulary and an unlimited creativity of speakers to combine words and to break the rules of grammar. Almost all existing automatic speech recognisers use statistics over speech sounds "what is the probability that a piece of audio is an a-sound" and statistics over word combinations to deal with this complexity. The ...

When exposed to noise, speakers will modify the way they speak in an effort to maintain intelligible communication. This process, which is referred to as Lombard effect (LE), involves a combination of both conscious and subconscious articulatory adjustment. Speech production variations due to LE can cause considerable degradation in automatic speech recognition (ASR) since they introduce a mismatch between parameters of the speech to be recognized and the ASR system’s acoustic models, which are usually trained on neutral speech. The main objective of this thesis is to analyze the impact of LE on speech production and to propose methods that increase ASR system performance in LE. All presented experiments were conducted on the Czech spoken language, yet, the proposed concepts are assumed applicable to other languages. The first part of the thesis focuses on the design and acquisition of a ...

Speech signals carry important information about a speaker such as age, gender, language, accent and emotional/psychological state. Automatic recognition of speaker characteristics has a wide range of commercial, medical and forensic applications such as interactive voice response systems, service customization, natural human-machine interaction, recognizing the type of pathology of speakers, and directing the forensic investigation process. This research aims to develop accurate methods and tools to identify different physical characteristics of the speakers. Due to the lack of required databases, among all characteristics of speakers, our experiments cover gender recognition, age estimation, language recognition and accent/dialect identification. However, similar approaches and techniques can be applied to identify other characteristics such as emotional/psychological state. For speaker characterization, we first convert variable-duration speech signals into fixed-dimensional vectors suitable for classification/regression algorithms. This is performed by fitting a probability density function to acoustic ...

A method is presented to improve speech recognition for pluricentric languages. Both the effect of adaptation of acoustic data and phonetic transcriptions for several subregions of the German speaking area are investigated and discussed. All experiments were carried out for German spoken in Germany and Austria using large telephone databases (Speech-Dat). In the first part triphone-based acoustic models (AMOs) were trained for several regions and their word error rates (WERs) were compared. The WERs vary between 9.89% and 21.78% and demonstrate the importance of regional variety adaptation. In the pronunciation modeling part narrow phonetic transcriptions for a subset of the Austrian database were carried out to derive pronunciation rules for Austrian German and to generate phonetic lexica for Austrian German which are the first of their kind. These lexica were used for both triphone-based and monophone-based AMOs with German and ...

The subject of this PhD thesis is the efficient and robust processing and analysis of the audio recordings that are derived from a call center. The thesis is comprised of two parts. The first part is dedicated to dialogue/non-dialogue detection and to speaker segmentation. The systems that are developed are prerequisite for detecting (i) the audio segments that actually contain a dialogue between the system and the call center customer and (ii) the change points between the system and the customer. This way the volume of the audio recordings that need to be processed is significantly reduced, while the system is automated. To detect the presence of a dialogue several systems are developed. This is the first effort found in the international literature that the audio channel is exclusively exploited. Also, it is the first time that the speaker utterance ...

This thesis introduces new learning algorithms for Radial Basis Function (RBF) networks. RBF networks is a feed-forward two-layer neural network used for functional approximation or pattern classification applications. The proposed training algorithms are based on robust statistics. Their theoretical performance has been assessed and compared with that of classical algorithms for training RBF networks. The applications of RBF networks described in this thesis consist of simultaneously modeling moving object segmentation and optical flow estimation in image sequences and 3-D image modeling and segmentation. A Bayesian classifier model is used for the representation of the image sequence and 3-D images. This employs an energy based description of the probability functions involved. The energy functions are represented by RBF networks whose inputs are various features drawn from the images and whose outputs are objects. The hidden units embed kernel functions. Each kernel ...

In a natural acoustic environment, multiple sources are usually active at the same time. The task of source separation is the estimation of individual source signals from this complex mixture. The challenge of single channel source separation (SCSS) is to recover more than one source from a single observation. Basically, SCSS can be divided in methods that try to mimic the human auditory system and model-based methods, which find a probabilistic representation of the individual sources and employ this prior knowledge for inference. This thesis presents several strategies for the separation of two speech utterances mixed into a single channel and is structured in four parts: The first part reviews factorial models in model-based SCSS and introduces the soft-binary mask for signal reconstruction. This mask shows improved performance compared to the soft and the binary masks in automatic speech recognition ...

Speech recognition performance degrades in the presence of background noise. In this thesis, several methods are developed to improve the noise robustness. Most of the work pertains to the use of sparse representations of speech: speech segments are described as a sparse linear combination of example speech segments, exemplars. Using techniques from missing data theory and compressed sensing, it is proposed to find, for each noisy speech observation, a sparse linear combination of exemplars using only speech features that are not corrupted by noise. This linear combination of clean speech exemplars is then used to reconstruct and estimate of the clean speech. Later in the thesis, it is proposed to augment this model by expressing noisy speech as a linear combination of speech and noise exemplars. Additionally, the weights of labelled exemplars in the sparse representation is used directly for ...

This thesis addresses the problem of multichannel audio source separation by exploiting deep neural networks (DNNs). We build upon the classical expectation-maximization (EM) based source separation framework employing a multichannel Gaussian model, in which the sources are characterized by their power spectral densities and their source spatial covariance matrices. We explore and optimize the use of DNNs for estimating these spectral and spatial parameters. Employing the estimated source parameters, we then derive a time-varying multichannel Wiener filter for the separation of each source. We extensively study the impact of various design choices for the spectral and spatial DNNs. We consider different cost functions, time-frequency representations, architectures, and training data sizes. Those cost functions notably include a newly proposed task-oriented signal-to-distortion ratio cost function for spectral DNNs. Furthermore, we present a weighted spatial parameter estimation formula, which generalizes the corresponding exact ...