The Effect Of Acoustic Variability On Automatic Speaker Recognition Systems

This thesis examines the influence of acoustic variability on automatic speaker recognition systems (ASRs) with three aims. i. To measure ASR performance under 5 commonly encountered acoustic conditions; ii. To contribute towards ASR system development with the provision of new research data; iii. To assess ASR suitability for forensic speaker comparison (FSC) application and investigative/pre-forensic use. The thesis begins with a literature review and explanation of relevant technical terms. Five categories of research experiments then examine ASR performance, reflective of conditions influencing speech quantity (inhibitors) and speech quality (contaminants), acknowledging quality often influences quantity. Experiments pertain to: net speech duration, signal to noise ratio (SNR), reverberation, frequency bandwidth and transcoding (codecs). The ASR system is placed under scrutiny with examination of settings and optimum conditions (e.g. matched/unmatched test audio and speaker models). Output is examined in relation to baseline performance and metrics assist in informing if ASRs should be applied to suboptimal audio recordings. Results indicate that modern ASRs are relatively resilient to low and moderate levels of the acoustic contaminants and inhibitors examined, whilst remaining sensitive to higher levels. The thesis provides discussion on issues such as the complexity and fragility of the speech signal path, speaker variability, difficulty in measuring conditions and mitigation (thresholds and settings). The application of ASRs to casework is discussed with recommendations, acknowledging the different modes of operation (e.g. investigative usage) and current UK limitations regarding presenting ASR output as evidence in criminal trials. In summary, and in the context of acoustic variability, the thesis recommends that ASRs could be applied to pre-forensic cases, accepting extraneous issues endure which require governance such as validation of method (ASR standardisation) and population data selection. However, ASRs remain unsuitable for broad forensic application with many acoustic conditions causing irrecoverable speech data loss contributing to high error rates.