Query-by-example spoken document retrieval (QbESDR) is a task that consists in retrieving those documents where a given spoken query appears. Spoken documents and queries exhibit a huge variability in terms of speaker, gender, accent or recording channel, among others. According to previous work, reducing this variability when following zero-resource QbESDR approaches, where acoustic features are used to represent the documents and queries, leads to improved performance. This work aims at reducing gender variability using voice conversion (VC) techniques. Specifically, a target gender is selected, and those documents and queries spoken by speakers of the opposite gender are converted in order to make them sound like the target gender. VC includes a resynthesis stage that can cause distortions in the resulting speech so, in order to avoid this, the use of the converted Mel-cepstral coefficients obtained from the VC system is proposed for QbESDR instead of extracting acoustic features from the converted utterances. Experiments were run on a QbESDR dataset in Basque language, and the results showed that the proposed gender variability reduction technique led to a relative improvement by 17% with respect to using the original recordings.
Cite as: López Otero, P., Docío-Fernández, L. (2018) Converted Mel-Cepstral Coefficients for Gender Variability Reduction in Query-by-Example Spoken Document Retrieval. Proc. IberSPEECH 2018, 82-86, doi: 10.21437/IberSPEECH.2018-18
@inproceedings{lopezotero18_iberspeech,
author={Paula {López Otero} and Laura Docío-Fernández},
title={{Converted Mel-Cepstral Coefficients for Gender Variability Reduction in Query-by-Example Spoken Document Retrieval}},
year=2018,
booktitle={Proc. IberSPEECH 2018},
pages={82--86},
doi={10.21437/IberSPEECH.2018-18}
}