Towards multiple vocal effort speaker verification: \nexploring speaker-dependent invariant information \nbetween normal and whispered speech.
Year: 2016
Type: dissertation
Abstract: Speech-based biometrics is one of the most effective ways for identity management and one of the \npreferred methods by users and companies given its flexibility, speed and reduced cost. In speaker \nrecognition, the two most popular tasks are speaker identification (SI) and speaker verification (SV). \nCommonly, SV exhibits greater practical applications related to SI, especially in access control and \nidentity management applications. Current state-of-the-art SV systems are known to be strongly \ndependent on the condition of the speech material provided as input and can be affected by unexpected \nvariability presented during testing, such as environmental noise or changes in vocal effort. \nIn this thesis, SV using whispered speech is explored, as whispered speech is known to be a natural \nspeaking style that despite its reduced perceptibility, still contains relevant information regarding \nthe intended message (i.e., intelligibility), as well as the speaker identity and gender. However, \ngiven the acoustic differences between whispered and normally-phonated speech, speech applications \ntrained on the latter but tested with the former exhibit unacceptable performance levels. Within \nan automated speaker verification task, previous research has shown that i) conventional features \n(e.g., mel-frequency cepstral coefficients, MFCCs) do not convey sufficient speaker discrimination \ncues across the two vocal efforts, and ii) multi-style training, while improving the performance for \nwhispered speech, tends to deteriorate the performance for normal speech. Moreover, by exploring \nthe performance boundaries achievable with whispered speech for speaker verification, it was found \nthat the lack of sufficient data to train whispered speech speaker models is a major limiting factor. \nAs such, this thesis aims to address these three shortcomings and proposes i) innovative features \nthat are less sensitive to changes from normal to whispered speech, thus helping to reduce the mismatch \ngap, ii) fusion strategies that improve accuracy for both normal and whispered speech, and \niii) machine learning principles that overcome the limited resources/data problem. \nTo properly address the task at hand, we first perform a comparative analysis among some \nof the most common feature based approaches and training techniques reported in related fields \nthat have shown benefits in the mismatch problem induced by vocal effort variations, including \nwhispered speech. This allowed us to narrow down candidate solutions and strategies that could be \nuseful for whispered speech speaker verification after some careful fine tuning. Next, by combining \nthese insights, with those previously reported in the literature, as well as statistical analysis tools, \ninnovative features are proposed that provide not only invariant information across the two vocal \nefforts, but also features that provide complementary information. Subsequently, in order to take \nadvantage of the complementary information extracted from the different feature representations, \nwe explore fusion schemes at different levels. When using a combination of normal and whispered \nspeech data during parameter estimation, improved multi vocal effort speaker verification is achieved \nand relative improvements as high as 68% and 70% for normal and whispered speech, respectively, \ncould be seen relative to a baseline system based on MFCC features. When including whispered \nspeech during enrollment, further improvements are achieved for whispered speech without severely affecting performance for normal speech. These results show that the proposed strategies allow to \nefficiently use the limited resources available, and achieve high performance levels for whispered \nspeech inline with performance obtained for normal speech.
Author: Sarria Paja, Milton Orlando
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7043713877
