Problems of Forensic Sciences 2013 Vol. 95 (XCV) 642-650
NOISE REDUCTION IN AUDIO RECORDINGS FOR FORENSIC PURPOSES USING AN ALGORITHM FOR SHORT-TIME SPECTRAL ANALYSIS
Marcin MICHAŁEK
Institute of Forensic Research, Kraków, Poland
Streszczenie
Forensic speech to text transcriptions from evidence recordings, which are usually intensively distorted, are carried out in almost every forensic audio analysis. This is connected with a need to correct recordings to improve the quality of the speech signal (forensic audio enhancement). The purpose of this work was to develop our own software to reduce various kinds of noises contained in evidence audio recordings, enabling improvement of speech signal intelligibility. During project implementation, an algorithm for adaptive noise reduction was used with short-time spectral amplitude estimation (MMSE STSA). Analyses conducted using the constructed software allowed us to evaluate the effectiveness of the reconstruction of the speech signal for different kinds of noises and SNR values, and define the threshold below which correct transcription is difficult or impossible. The results made it possible to indicate what kind of noise has the most destructive effect on the speech signal and what conditions must be fulfilled by recordings in order to enable a correct transcription. The impact of the applied algorithm on possible distortions of the speech signal after reconstruction was also estimated. The designed and validated software for noise reduction will be a useful tool that allows more effective audio analysis in forensic practice.
Słowa kluczowe
Forensic audio analysis; Audio recording; Noise cancellation; Spectral analysis.