Effects of emotional speech on forensic voice comparison using deep speaker embeddings

Abed Mohammed Hamzah; Sztahó Dávid: Effects of emotional speech on forensic voice comparison using deep speaker embeddings.

Előnézet

Cikk, tanulmány, mű msznykonf_019_159-170..pdf Letöltés (1MB) | Előnézet

Absztrakt (kivonat)

Emotional conditions play a significant role in forensic voice comparison and speaker verification systems. When emotion is present in speech, the verification's performance will deteriorate. In this paper, speaker verification has been investigated and analyzed in the case of emotional speech using metrics evaluating the performance of forensic voice comparison using pre-trained speaker embedding models: x-vector and ECAPA-TDNN for embedded feature extraction. This study investigates whether emotional content affects the forensic voice comparison and verification performance evaluated on a Hungarian speech dataset. The speaker verification performance has been assessed using the likelihood-ratio framework using Cllr and Cllrmin and Equal Error Rate. The ECAPATDNN achieved higher performance than the x-vector. In the same emotion scenario, the best EERs were 2.6% and 7.7% for ECAPA-TDNN and x-vector. Both models are sensitive to the emotional content of the speech samples.

Tétel nézet