On the Influence of Automatic Segmentation and Clustering in Automatic Speech Recognition

An automatic speech recognition (ASR) system needs a previous segmentation stage that differentiates between speech and non-speech. Other information such as "who spoke when" can be proportioned to the ASR system, allowing it to perform speaker adaptation. This paper studies the influence of automatic speech segmentation and speaker clustering on ASR performance, in order to detect the weak points of the diarization system by analyzing what causes the different types of recognition errors: insertions, suppressions and substitutions. Experiments are run on the Galician broadcast news database Transcrigal, and results show that the speaker diarization system presented in this work is suitable as a previous step to ASR, as the performance is almost the same as the obtained when using manual segmentation and clustering.