Vowel based Speaker Verification using Self Organising Map

Speaker verification using heterogeneous neural network architecture with linear correlation speech activity detection

“…In previous work, the authors compared the performance of the three main feature formats used in speaker recognition systems (DFT, LPC and MFCC) when used as input feature vectors to a SOM together with a rule‐based scoring speaker verification system (Tashan et al ., ). The results show that the DFT spectrum is better than the MFCC and the LPC spectrum features with the system adopted.…”

“…The first stage is a frame filtering stage that uses the modified SOM, presented by Tashan et al . () (and described in Section 3.1), as a claimed user voice model for the three vowels considered in the experiment. The second stage then consists of three MLP networks (described in Section 3.2); each of which has been trained to function as a claimed user vowel verifier.…”

“…Most modern voice biometric authentication systems employ GMM‐based methods in the verification engine, an offshoot of earlier research into the use of HMM algorithms for speech recognition systems. SOM‐based speaker recognition systems, on the other hand, are attractive alternatives to the conventional methods because they offer the potential of being able to do away with the need for the speech recognition front‐end, commonly included in speaker recognition systems (Tashan et al ., ).…”

“…Although the MFCC are the most preferred input feature format in the literature, there is evidence to suggest that these may not be optimal for neural network‐based systems (Sun et al ., ) (Tashan et al ., ). A recent experiment was conducted to explore the use of the three feature formats mentioned previously as input feature vector for a SOM‐based speaker verification system (Tashan et al ., ). This experiment shows that the DFT spectrum performs better than the MFCC and the LPC spectrum features with the system adopted.…”

“…Pre‐processing of the input speech samples is achieved using either energy frame analysis (EFA)‐based speech activity detection or a new DFT spectrum‐based linear correlation technique. The first speaker verification stage consists of a modified SOM (Tashan et al ., ) that is used to filter the speech signal to select only those speech frames that are similar to enrolled vowel phoneme frames for the claimed speaker. The linear correlation pre‐processing, in combination with the SOM filtering stage, effectively replaces the speech recognition tool used in traditional speaker recognition systems.…”