A Semi-supervised Speaker Identification Method for Audio Forensics Using Cochleagrams
| Autor: | Steven Camacho, M L Dora Ballesteros, Diego Renza |
|---|---|
| Rok vydání: | 2017 |
| Předmět: | |
| Zdroj: | Communications in Computer and Information Science ISBN: 9783319669625 WEA |
| DOI: | 10.1007/978-3-319-66963-2_6 |
| Popis: | The general task in speaker identification for audio forensics is to identify the unknown speaker within an audio proof, who is suspected of a crime. Here, the voice of each person within a group of suspects is compared to the audio proof with the aim to determining which of them corresponds to the source. In this paper, a semi-supervised speaker identification method is proposed, which does not require a training stage. Also, the feature extraction is based on the use of cochleagrams for the previously selected words. The system can identify one or multiple suspects which have high similarity to the audio proof, or give a null response if none of the suspects satisfies a similarity threshold. The results of the proposed method are compared with the respective results of the same method but using spectrograms instead of cochleagrams. The performance of our system is measured through a confusion matrix (true and false positives, and true and false negatives) and global results are given in terms of overall accuracy and kappa index. According to several tests, our system has an overall accuracy higher than 0.97 and a kappa index around 0.78; this means a high confidence in the results of identification and rejection. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|