Retrieval and browsing of spoken content
IEEE Signal Processing Magazine, cilt.25, sa.3, ss.39-49, 2008 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 25 Sayı: 3
- Basım Tarihi: 2008
- Doi Numarası: 10.1109/msp.2008.917992
- Dergi Adı: IEEE Signal Processing Magazine
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
- Sayfa Sayıları: ss.39-49
- Boğaziçi Üniversitesi Adresli: Evet
Özet
The development of information retrieval systems for spoken audio documents have challenges regarding the issue of handling the errorful or incomplete output provided by automatic speech recognition (ASR) systems. Nevertheless, such systems work by permitting user to enter search terms into a special search engine and then producing a collection of returned spoken documents called hits. Despite word-error-rates of as high as 50%, experimental evidence shows that using the spoken content in an audio collection provides a significant improvement in performance with respect to using only accompanying text metadata for retrieval. However, spoken document retrieval (SDR) produces challenges for ASR technology especially for large and non homogeneous document collections where tuning an ASR system to narrow domains is not a scalable approach. Nevertheless, there are approaches on mitigating ASR limitations, such as soft-indexing, which permit the precision-recall operating point to be adjusted to the needs of a given application.