In recent years, there has been an increasing interest in the field of biometric recognition centred around face, fingerprint and iris, as the most reliable modalities. While speaker recognition holds considerable promise, its widespread adoption has been hindered by suboptimal performance marked by low accuracy and reliability issues. However, addressing these challenges is essential to fully exploit the capabilities of voice recognition, particularly given surveys indicating a strong customer preference for use of this modality. In situations where remote speaker biometric authentication is employed, maintaining the integrity of extracted cepstral feature vectors becomes paramount in the presence of channel noise. To enhance the accuracy of speaker recognition in remote applications, a novel approach is proposed where Low-Density Parity-Check (LDPC) codes are used for both storage and transmission of extracted Mel-frequency cepstral coefficients (MFCC) derived from voice samples from a limited database. Utilizing a code rate of ½ and a block length of 1024 bits, an examination of the bit error rate in the presence of channel noise indicates a minimal error floor. The integration of LDPC decoded MFCC features led to a notable enhancement in accuracy within the speaker recognition systems, surpassing the performance of the standalone system relying on uncoded features. The methods proposed in this paper, upon testing on locally created database, exhibit an average verification time of 4 s. The low error rate and high accuracy rate results are suggestive to the use of LDPC coding as an integral part of remote speaker recognition procedure.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Augmenting Reliability in Speaker Recognition Systems Through Low Density Parity Check Codes

  • Nilashree S. Wankhede,
  • Sushama Wagh

摘要

In recent years, there has been an increasing interest in the field of biometric recognition centred around face, fingerprint and iris, as the most reliable modalities. While speaker recognition holds considerable promise, its widespread adoption has been hindered by suboptimal performance marked by low accuracy and reliability issues. However, addressing these challenges is essential to fully exploit the capabilities of voice recognition, particularly given surveys indicating a strong customer preference for use of this modality. In situations where remote speaker biometric authentication is employed, maintaining the integrity of extracted cepstral feature vectors becomes paramount in the presence of channel noise. To enhance the accuracy of speaker recognition in remote applications, a novel approach is proposed where Low-Density Parity-Check (LDPC) codes are used for both storage and transmission of extracted Mel-frequency cepstral coefficients (MFCC) derived from voice samples from a limited database. Utilizing a code rate of ½ and a block length of 1024 bits, an examination of the bit error rate in the presence of channel noise indicates a minimal error floor. The integration of LDPC decoded MFCC features led to a notable enhancement in accuracy within the speaker recognition systems, surpassing the performance of the standalone system relying on uncoded features. The methods proposed in this paper, upon testing on locally created database, exhibit an average verification time of 4 s. The low error rate and high accuracy rate results are suggestive to the use of LDPC coding as an integral part of remote speaker recognition procedure.