Breathprint: Speaker Recognition Based On Breathing

The identification of the Speaker through Speech Recognition has widely been researched and used in our daily lives. It is employed in automatic telephone services, in the field of audio forensics and so on. However, the prior research mainly focuses on text-dependent based systems, which greatly restrict this application for the identification of speakers by using limited and specific speech contents. In contrast, the proposed research is aimed at expanding the speaker recognition system to a text-independent system in which the speech contents are not limited. Within the realm of text-independent speaker identification, the chief research aspects include: spectrum characteristics, voice-source characteristics, rhythm and vocabulary or more high-level features. Even still, the challenge remains to extract the requisite features and improve the identification accuracy and dependency on speech contents. Besides, high level features result to complex and computational algorithms, which is yet another concern for a processing-friendly and real-time system.We focus on exploring a novel speaker recognition scheme which is simple, effective and efficient. For the first time, we propose a speaker recognition scheme based on breath of the speaker instead of speech contents. Out research stems from the observation that the breath segments are unaffected by diversity of speech contents and mainly rely on biological features of vocal parts(like, lungs, mouth, trachea etc). The main idea is to extract and process the breath contents in speech to devise a model for each speaker, and then use this model to identify different speakers.In this thesis, we made a complete breath extraction, modeling and speaker recognition system and tested it against 34 volunteers and 340 breath samples. By using the MFCC processing of breath segments, our scheme successfully identifies the speaker with both FAR and FRR below 10%. The accuracy of our algorithm improves for longer speech signals with FAR and FRR below 5%.