Please Speaker Recognition Phd Thesis note. All our papers are written from scratch. To ensure high quality of writing, the pages number is limited for short deadlines. If you want to order more pages, please choose longer Deadline (Urgency) Recently Rated. + (Only For WhatsApp) + (Phone Number) Free. Have routine homework and. Speaker Recognition Phd Thesis. academic assignments completed at affordable prices. Give us your assignments and a subject matter expert will get it done quickly and painlessly. Better grades can be yours without stress! Speaker Recognition Phd Thesis to the required academic referencing style, such as APA, MLA, Harvard and Chicago. Thus, being written and edited by our professionals, your essay will achieve perfection. Our Speaker Recognition Phd Thesis writing staff is working to meet your needs and expectations and take care of your writing assignment!
edu no longer supports Internet Explorer. To browse Academia. edu and the wider internet faster and more securely, please take a few seconds to upgrade your browser. Log In with Facebook Log In with Google Sign Up with Apple. Remember me on this computer. Enter the email address you signed up with and we'll email you a reset link. Need an account? Click here to sign up. Download Free PDF. On the use of PCA in GMM and AR-vector models for text independent speaker verification 14th International Conference on Digital Signal Processing Proceedings.
DSP Cat. Apolinario Jr. Download PDF Download Full PDF Package This paper. A short summary of this paper. On the use of PCA in GMM and AR-vector models for text independent speaker verification. Apolinário Jr. br Abstract: This paper examines the role of the Principal Components Analysis PCA on the performance of two classification systems for text independent speaker verification: the Gaussian Mixture Model GMM and the AR-Vector Model.
The use of the PCA transform resulted in an improvement in the performance of the GMM for training times of 60s and 30s. However, the advantage of using PCA was not observed for the AR-Vector model. For the case of 10s training time, there was no benefit in using PCA even with GMM. In this situation, the AR-Vector is superior for a 10s test and worse for a 3s test.
Speech, being present everywhere from telephone nets to The Principal Components Analysis PCA is used with personal computers, may be the cheapest form to supply a the purpose of decorrelating the training data.
This leads growing need of providing authenticity and privacy in the to easier statistical models and adds structural information worldwide communication nets [1]. Speaker verification from the training data eigenvalues of the covariance ma- is the task of verifying if a speech signal utterance be- trix in an effort to provide more discriminative features to longs or not to a certain person, which means a binary deci- the speaker recognition system, speaker recognition phd thesis.
The decisions are carried out in the so-called speakers This paper is organized as follows. Section 2 and 3 open set [2] because the recognition is done in an unknown briefly review the GMM and the AR-Vector, respectively. speakers set possible impostors. As to text dependency, The PCA is presented in Section 4. Section 5 contains de- recognition can be dependent or independent. Systems de- tails of the system setup and presents the simulation results.
manding a predetermined word or phrase are text depen- Concluding remarks are given in Section 6. Two classification systems using PCA are investigated in this paper: the GMM and the AR-Vector.
The GMM [3] combines the robustness and smoothing properties of the 2. THE GAUSSIAN MIXTURE MODEL parametric Gaussian model with the arbitrary modeling ca- pability speaker recognition phd thesis a non-parametric VQ.
vast phonetic class to characterize the sound produced by a Each component density is a D variate Gaussian func- person [4]. In speaker recognition 1; ; M. A k termined in order to maximize the likelihood speaker recognition phd thesis the GMM.
k are obtained by solving the following set of equa- The algorithm presented in [3] is widely used for this tions: task. The scale factor 1 is used in order B R 1 R 2 R0 Speaker recognition phd thesis R T T to normalize the likelihood according to the duration of the p p p p utterance number of feature vectors. The use of the Itakura distance with the AR-Vector is presented in [5].
Assuming a stored model A previously estimated from a given speaker Fig, speaker recognition phd thesis. Speaker verification system using GMM, speaker recognition phd thesis.
and a model B from a pretense speaker, three distance mea- sures between these two models are defined for their respec- Speaker recognition phd thesis system uses two models which provide the normal- tive autocorrelation matrices. The background is built The speaker verification system provides a binary output, ac- with a hypothetical set of false speakers and modeled via ceptance or rejection of a pretense speaker.
Hence, an estimation GMM universal background model [6]. This threshold is estimated with the true distances, i. models under comparison are from the same person, speaker recognition phd thesis, and with the false distances given by the pretense speaker model compared to the other models not belonging to him.
THE AR-VECTOR MODEL From these distances, the speaker recognition phd thesis is estimated taking into account false acceptance errors and false rejection errors. When The AR-Vector is actually an extension of the LPC in the a speaker is to be analyzed, he or she will be accepted if the sense that it carries out a prediction among vectors not resulting distance is lower than the threshold.
He or she will samplesmodeling the time evolving of the vectors in our be rejected otherwise. The order p AR-Vector system. X nique widely used in pattern recognition [7]. Each speaker uttered sentences, in Brazilian Portuguese, extracted from [9]. The silence between words were eliminated.
The number of Gaussians for the GMM was set to 32 while AR-Vector Itakura Distance d used order 2 with the symmetric Itakura distance previous exper- iments have shown its better performance for this configuration. We have used 60, 30, and 10s of speech signal for training and 30, 10, speaker recognition phd thesis, and 3s for testing. This procedure resulted in an equal error rate EER measure [2]. AR-Vector Speaker Verification System.
The results obtained with the 32 Gaussians GMM will be com- pared to the order 2 AR-Vector using symmetric Itakura distance, using both MCC and MCCPCA PCA transformed MCC vectors. We ear combination of its components. This than for the AR-Vector. Moreover, the performance of the GMM procedure also corresponds to the discrete version of the Karhunen- was better than the performance of the AR-Vector for 10s and 3s Loéve Transform. of testing time, mainly in the latter case.
The transformation performed by ma- GMM - MCC 0 0. AR-Vector - MCC AR-Vector - MCCPCA 0 0 1. Each speaker will have an associ- Table 2 presents the results for 30s of training. These results ated transformation matrix. Therefore, each of these transforma- show that the PCA technique still favors the GMM but speaker recognition phd thesis same is tion matrices—e. For 30s test, the AR-Vector has pre- covariance matrix of its correspondent speaker, speaker recognition phd thesis.
test phase. For 3s test, the performance of the AR-Vector is almost 3 times lower than the GMM. They resulted in the highest error rates of both classification sys- This section details the setup and the results of the speaker verifica- tems. When the training time is 10s, the use of the PCA presented tion system implemented in our experiments. The utterances were no significant improvement—and eventually loss of performance.
recorded with 8KHz as sampling rate, electret microphones, and The AR-Vector presented better results as compared to the GMM in a low noise environment. Performance of the GMM versus the AR-Vector, with and without PCA, for 30s training.
Proceedings of the IEEE, vol. GMM - MCC 1. Speaker Identification and Verifica- AR-Vector - MCC 0 1. Speech Com- AR-Vector - MCCPCA 0 1. A Gaussian Mixture Modeling Ap- Table 3. Performance of the GMM versus the AR-Vector, proach to Text Independent Speaker Identification. PhD The- sis. Georgia Institute of Technology, August with and without PCA, for 10s training.
Robust Text-Independent Speaker System 10s 3s Identification Using Gaussian Mixture Speaker Model. IEEE EER EER Transactions on Speech and Audio Processing, vol. GMM - MCCPCA 4. Mathan, speaker recognition phd thesis, A. de Lima, and G. Standard and Target Driven AR-vector Models for Speech Analysis and Speaker Recognition. Proceedings of ICASSP, San Francisco, USA, vol.
II5-II8, March Throughout the analysis of the results presented here, we can clearly note that the amount of time for training and for testing has [6] REYNOLDS, Douglas A. Thomas F. Quatieri, and Robert B.
a strong influence, speaker recognition phd thesis. The larger they are the more statistics they are Dunn.
[ICASSP 2018] Google's D-Vector System: Generalized End-to-End Loss for Speaker Verification
, time: 17:34
Speaker Recognition Phd Thesis to the required academic referencing style, such as APA, MLA, Harvard and Chicago. Thus, being written and edited by our professionals, your essay will achieve perfection. Our Speaker Recognition Phd Thesis writing staff is working to meet your needs and expectations and take care of your writing assignment! It's safer that way and helps avoid any uncomfortable Speaker Recognition Phd Thesis questions. You're the author and that's the way it goes. Turn it custom-written papers, get above-average grades, and still have plenty of time for hobbies, friends, parties, and Speaker Recognition Phd Thesis career/10() Recently Rated. + (Only For WhatsApp) + (Phone Number) Free. Have routine homework and. Speaker Recognition Phd Thesis. academic assignments completed at affordable prices. Give us your assignments and a subject matter expert will get it done quickly and painlessly. Better grades can be yours without stress!
No comments:
Post a Comment