Aquesta tesi doctoral mostra la recerca feta en l'àrea de la diarització de locutor per a sales de reunions. En la present s'estudien els algorismes i la implementació d'un sistema en diferit de segmentació i aglomerat de locutor per a grabacions de reunions a on normalment es té accés a més d'un micròfon per al processat. El bloc més important de recerca s'ha fet durant una estada al International Computer Science Institute (ICSI, Berkeley, Caligornia) per un període de dos anys.La diarització de locutor s'ha estudiat força per al domini de grabacions de ràdio i televisió. La majoria dels sistemes proposats utilitzen algun tipus d'aglomerat jeràrquic de les dades en grups acústics a on de bon principi no se sap el número de locutors òptim ni tampoc la seva identitat. Un mètode molt comunment utilitzat s'anomena "bottom-up clustering" (aglomerat de baix-a-dalt), amb el qual inicialment es defineixen molts grups acústics de dades que es van ajuntant de manera iterativa fins a obtenir el nombre òptim de grups tot i acomplint un criteri de parada. Tots aquests sistemes es basen en l'anàlisi d'un canal d'entrada individual, el qual no permet la seva aplicació directa per a reunions. A més a més, molts d'aquests algorisms necessiten entrenar models o afinar els parameters del sistema usant dades externes, el qual dificulta l'aplicabilitat d'aquests sistemes per a dades diferents de les usades per a l'adaptació.La implementació proposada en aquesta tesi es dirigeix a solventar els problemes mencionats anteriorment. Aquesta pren com a punt de partida el sistema existent al ICSI de diarització de locutor basat en l'aglomerat de "baix-a-dalt". Primer es processen els canals de grabació disponibles per a obtindre un sol canal d'audio de qualitat major, a més dínformació sobre la posició dels locutors existents. Aleshores s'implementa un sistema de detecció de veu/silenci que no requereix de cap entrenament previ, i processa els segments de veu resultant amb una versió millorada del sistema mono-canal de diarització de locutor. Aquest sistema ha estat modificat per a l'ús de l'informació de posició dels locutors (quan es tingui) i s'han adaptat i creat nous algorismes per a que el sistema obtingui tanta informació com sigui possible directament del senyal acustic, fent-lo menys depenent de les dades de desenvolupament. El sistema resultant és flexible i es pot usar en qualsevol tipus de sala de reunions pel que fa al nombre de micròfons o la seva posició. El sistema, a més, no requereix en absolute dades d´entrenament, sent més senzill adaptar-lo a diferents tipus de dades o dominis d'aplicació. Finalment, fa un pas endavant en l'ús de parametres que siguin mes robusts als canvis en les dades acústiques. Dos versions del sistema es van presentar amb resultats excel.lents a les evaluacions de RT05s i RT06s del NIST en transcripció rica per a reunions, a on aquests es van avaluar amb dades de dos subdominis diferents (conferencies i reunions). A més a més, es fan experiments utilitzant totes les dades disponibles de les evaluacions RT per a demostrar la viabilitat dels algorisms proposats en aquesta tasca. / This thesis shows research performed into the topic of speaker diarization for meeting rooms. It looks into the algorithms and the implementation of an offline speaker segmentation and clustering system for a meeting recording where usually more than one microphone is available. The main research and system implementation has been done while visiting the International Computes Science Institute (ICSI, Berkeley, California) for a period of two years. Speaker diarization is a well studied topic on the domain of broadcast news recordings. Most of the proposed systems involve some sort of hierarchical clustering of the data into clusters, where the optimum number of speakers of their identities are unknown a priory. A very commonly used method is called bottom-up clustering, where multiple initial clusters are iteratively merged until the optimum number of clusters is reached, according to some stopping criterion. Such systems are based on a single channel input, not allowing a direct application for the meetings domain. Although some efforts have been done to adapt such systems to multichannel data, at the start of this thesis no effective implementation had been proposed. Furthermore, many of these speaker diarization algorithms involve some sort of models training or parameter tuning using external data, which impedes its usability with data different from what they have been adapted to.The implementation proposed in this thesis works towards solving the aforementioned problems. Taking the existing hierarchical bottom-up mono-channel speaker diarization system from ICSI, it first uses a flexible acoustic beamforming to extract speaker location information and obtain a single enhanced signal from all available microphones. It then implements a train-free speech/non-speech detection on such signal and processes the resulting speech segments with an improved version of the mono-channel speaker diarization system. Such system has been modified to use speaker location information (then available) and several algorithms have been adapted or created new to adapt the system behavior to each particular recording by obtaining information directly from the acoustics, making it less dependent on the development data.The resulting system is flexible to any meetings room layout regarding the number of microphones and their placement. It is train-free making it easy to adapt to different sorts of data and domains of application. Finally, it takes a step forward into the use of parameters that are more robust to changes in the acoustic data. Two versions of the system were submitted with excellent results in RT05s and RT06s NIST Rich Transcription evaluations for meetings, where data from two different subdomains (lectures and conferences) was evaluated. Also, experiments using the RT datasets from all meetings evaluations were used to test the different proposed algorithms proving their suitability to the task.

The thesis discusses in detail some concepts of speech and prosody that can contribute to build a speech corpus for the speaker segmentation purpose. Moreover, the Elan multimedia annotator used for labeling is described. The theoretical part highlights some frequently used speech features such as MFCC, PLP and LPC and deals with currently most popular speech segmentation methods. Some classification algorithms are also mentioned. The practical part describes implementation of Bayesian information criterium algorithm in system for automatic speaker segmentation. For classification of speaker change point in speech, were used different speech features. The results of tests were evaluated by the graphic method of receiver operating characteristic (ROC) and his quantitative indices. As the best speech features for this system were provided MFCC and HFCC.

This thesis concerns with the recognition of music instruments from isolated notes. Music instrument recognition is a relatively nascent problem fast gaining importance not only because of the academic value the problem provides, but also for the potential it has in being able to realize applications like music content analysis, music transcription etc. Line spectral frequencies are proposed as features for music instrument recognition and shown to perform better than Mel filtered cepstral coefficients and linear prediction cepstral coefficients. Assuming a linear model of sound production, features based on the prediction residual, which represents the excitation signal, is proposed. Four improvements are proposed for classification using Gaussian mixture model (GMM) based classifiers. One of them involves characterizing the regions of overlap between classes in the feature space to improve classification. Applications to music instrument recognition and speaker recognition are shown. An experiment is proposed for discovering the hierarchy in music instrument in a data-driven manner. The hierarchy thus discovered closely corresponds to the hierarchy defined by musicians and experts and therefore shows that the feature space has successfully captured the required features for music instrument characterization.

This work aims at a task of speaker diarization. The goal is to implement a system which is able to decide "who spoke when". Particular components of implementation are described. The main parts are feature extraction, voice activity detection, speaker segmentation and clustering and finally also postprocessing. This work also contains results of implemented system on test data including a description of evaluation. The test data comes from the NIST RT Evaluation 2005 - 2007 and the lowest error rate for this dataset is 18.52% DER. Results are compared with diarization system implemented by Marijn Huijbregts from The Netherlands, who worked on the same data in 2009 and reached 12.91% DER.