Memory of AMR coded speech distorted by packet loss
2014 (English)Conference paper (Refereed)Text
Previous studies have shown that free recall of spoken word lists is impaired if the speech is presented in background noise, even if the signal-to-noise ratio is kept at a level allowing full word identification. The objective of this study was to examine recall rates for word lists presented in noise and word lists coded by an AMR (Adaptive Multi Rate) telephone codec distorted by packet loss. Twenty subjects performed a word recall test. Word lists consisting of ten words were played to the subjects. The subjects repeated each word immediately after it had been played, to ensure that the words were heard correctly. After the complete list had been played the subjects wrote down all words remembered. In this way, both word identification and recall rates were measured. Three distorted conditions were compared with an undistorted control condition using a within-subject design: speech spectrum weighted noise at 4 dB SNR, and AMR coded speech with two levels of packet loss, one mild and one severe. The results confirmed the disruptive effect of noise on free recall of words, while no significant impairment was found for the AMR distortions. The noise and the AMR coding with mild packet loss gave approximately the same impairment of word identification. The AMR coding with severe packet loss gave a larger impairment of word identification, even though the word recall rate was unaffected. This result suggests that packet loss in AMR coded speech causes distortions which disrupt recall of words less than noise at levels resulting in the same change of word identification rates. Since impairment of word identification rates did not correlate with impairment of word recall rates models for quality prediction of speech reproductions should not be based on identification rates alone.
Place, publisher, year, edition, pages
European Acoustics Association (EAA), 2014. 349- p.
, Proceedings of Forum Acusticum, ISSN 22213767
Packet loss; Speech; Speech recognition, Adaptive multi rates; Background noise; Disruptive effects; Identification rates; Quality prediction; Speech spectra; Spoken words; Word identification, Signal to noise ratio
IdentifiersURN: urn:nbn:se:hig:diva-21070ScopusID: 2-s2.0-84953300359ISBN: 978-836140228-2OAI: oai:DiVA.org:hig-21070DiVA: diva2:894691
7th Forum Acusticum (FA 2014), 7-12 September 2014, Krakow, Poland