Metode și tehnologii de inteligență artificială aplicate în tehnologia vorbirii / Methods and technologies of artificial intelligence applied in speech technology

The main objective of the current work is to use artificial intelligence methods and technologies in order to bring improvements in 3 tasks from the speech technology field: automatic speech recognition (ASR), automatic annotation of audio corpora and automatic speaker recognition. The first two tasks were explored in depth, while the third task still has plenty of room for exploration. The greatest efforts and the most contributions were made in the areas of speech automatic annotation, as well as in the field of automatic speech recognition, with an emphasis on training ASR systems for the Romanian language. These two tasks are interdependent: the evolution of one of the directions also attracted the evolution of the other. This thesis presents the successive steps taken to train high-performance ASR systems, in parallel with the automatic annotation of new audio data sets, where the target language was Romanian. To the best of our knowledge, the final ASR system obtained results very close to the state of the art, having an accuracy of 99% on read speech, respectively 90%-95% on spontaneous speech, depending on the difficulty of the transcription task.