"Beethoven's Ninth" - An Experiment on Naming Usage for Audio Files

Authors: Sandra Mann, André Berton, Marcel Dausend, Paul Heisterkamp

Abstract:

Evidence from experimental data has shown that users have difficulties in properly remembering music and audio book titles according to categories such as artist/author, afbum and title (metadata). When it comes to providing in-car search for audio files it is therefore not sufficient to generate speakable ASR (Automatic Speech Recognition) vocabulary from the original metadata only. Users are very likely to come up with wording variants that might strongly deviate from the original lexicon entry. In this paper we present an approach that enables the system to retrieve audio files based on incomplete names. We pre-process audio file names, applying filter and recombination rules. In order to verify in how far our preprocessed metadata cover what users say when searching for audio files, we devised and conducted an experiment to collect user speech data. Each subject had to go through three different scenarios: completely unrestricted naming for music selection, naming of individual favourite music according to given categories, and recall and reproduction of given names and categories. In the latter case we provoked cognitive overload by presenting audio file names in pairs of two for a short time only. This fead the subjects to filter out crucial parts of the names presented, as we intended. A first analysis of the speech data is supporting our approach: users often address music by variants of the “official” name. Our first set of filter and recombination rules, derived from introspection and common sense seems to cover a majority of spoken variants guite well. A more fine-grained analysis of (he naming data will increase this coverage. We present the experimental setup, our analysis of the collected data and the results in more detail in the paper. We also give an outlook on future working directions.