User Specific Training of a Music Search Engine

Query-by-Humming (QBH) systems transcribe a sung or hummed query and search for related musical themes in a database, returning the most similar themes as a play list. A major obstacle to effective QBH is variation between user queries and the melodic targets used as database search keys. Since it is not possible to predict all individual singer profiles before system deployment, a robust QBH system should be able to adapt to different singers after deployment. Currently deployed systems do not have this capability. We describe a new QBH system that learns from user provided feedback on the search results, letting the system improve while deployed, after only a few queries. This is made possible by a trainable note segmentation system, an easily parameterized singer error model and a straight-forward genetic algorithm. Results show significant improvement in performance given only ten example queries from a particular user.

[1]  Alan Fern,et al.  Gradient Boosting for Sequence Alignment , 2006, AAAI.

[2]  William P. Birmingham,et al.  A Comprehensive Trainable Error Model for Sung Music Queries , 2004, J. Artif. Intell. Res..

[3]  Elaine Chew,et al.  A statistical approach to retrieval under user-dependent uncertainty in query-by-humming systems , 2004, MIR '04.

[4]  David A. Shamma,et al.  Teaching a Music Search Engine Through Play , 2007, CHI 2007.

[5]  Jim R. Parker.  Genetic Algorithms for Continuous Problems , 2002, Canadian Conference on AI.

[6]  Remco C. Veltkamp,et al.  A Survey of Music Information Retrieval Systems , 2005, ISMIR.

[7]  Emmanuel Vincent,et al.  The 2005 Music Information retrieval Evaluation Exchange (MIREX 2005): Preliminary Overview , 2005, ISMIR.

[8]  Steven W. Zucker,et al.  Local Scale Control for Edge Detection and Blur Estimation , 1996, IEEE Trans. Pattern Anal. Mach. Intell..

[9]  Sungdo Ha,et al.  Computer-Human Interaction, 8th Asia-Pacific Conference, APCHI 2008, Seoul, Korea, July 6-9, 2008, Proceedings , 2008, Asia-Pacific Computer and Human Interaction.

[10]  William P. Birmingham,et al.  Name that tune: A pilot study in finding a melody from a sung query , 2004, J. Assoc. Inf. Sci. Technol..

[11]  William P. Birmingham,et al.  HMM-based musical query retrieval , 2002, JCDL '02.

[12]  R. Shepard.  Geometrical approximations to the structure of musical pitch. , 1982, Psychological review.

[13]  George Tzanetakis,et al.  A comparative evaluation of search techniques for query-by-humming using the MUSART testbed , 2007 .

[14]  Tae-Sun Choi,et al.  Local threshold and Boolean function based edge detection , 1999, 1999 Digest of Technical Papers. International Conference on Consumer Electronics (Cat. No.99CH36277).

[15]  Peter Meer,et al.  Edge Detection with Embedded Confidence , 2001, IEEE Trans. Pattern Anal. Mach. Intell..

[16]  George Tzanetakis,et al.  A framework for audio analysis based on classification and temporal segmentation , 1999, Proceedings 25th EUROMICRO Conference. Informatics: Theory and Practice for the New Millennium.

[17]  Alden H. Wright,et al.  Genetic Algorithms for Real Parameter Optimization , 1990, FOGA.

[18]  Peter Cariani,et al.  Temporal Codes, Timing Nets, and Music Perception , 2001 .

[19]  Yasushi Sakurai,et al.  SoundCompass: a practical query-by-humming system; normalization of scalable and shiftable time-series data and effective subsequence generation , 2004, SIGMOD '04.

[20]  William P. Birmingham,et al.  The dangers of parsimony in query-by-humming applications , 2003, ISMIR.