MIR at Google: Strategies for Scaling to Large Music Datasets Using Ranking and Auditory Sparse-Code Representations
Douglas Eck (Google) (Invited speaker) – There’s no paper associated with this talk.

Machine Listening / Audio analysis – Dick Lyon and Samy Bengio

Main strength:

• Scalable algorithms
  • When they do work, they use large sets (like all audio on Youtube, or all audio on the web)
• Sparse High dimensional Representations
  • 15 numbers to describe a track
• Auditory / Cohchlear Modeling
• Autotagging at Youtube –
• Retrieval, annotation, ranking, recommendation

Collaboration Opportunities

• Faculty research awards
• Google visiting faculty program
• Student internships
• Google summer of code
• Research Infrastructure

The Future of MIR is already here

• Next generation of listeners are using Youtube – because of the on-demand nature
• Youtube – 2 billion views a day
• Content ID scans over 100 years of video every day

The Bar is already set very high ..

• Current online recommendation is pretty good
• Doug wants to close the loop between music making and music listening

What would you like Google to give back to MIR?

A Roadmap Towards Versatile MIR
Emmanuel Vincent, Stanislaw A. Raczyński, Nobutaka Ono and Shigeki Sagayama

ABSTRACT – Most MIR systems are specifically designed for one appli- cation and one cultural context and suffer from the seman- tic gap between the data and the application. Advances in the theory of Bayesian language and information process- ing enable the vision of a versatile, meaningful and accu- rate MIR system integrating all levels of information. We propose a roadmap to collectively achieve this vision.

Wants to increase versatility of MIR systems across different types of music.  Systems adopt a fixed expert viewpoint ( musicologist, musician).  Have limited accuracy due to general pattern recognition techniques applied to a bag of features.

Emannuel wants to build an overarching scalable MIR system that successfully deals with the challenge on scalable unsupervised methods and refocuses MIR on symbolic methods.   This is the core roadmap of VERSAMUS.

The aim of VERSAMUS is to investigate, design and validate such representations in the framework of Bayesian data analysis, which provides a rigorous way of combining separate feature models in a modular fashion. Tasks to be addressed include the design of a versatile model structure, of a library of feature models and of efficient algorithms for parameter inference and model selection. Efforts will also be dedicated towards the development of a shared modular software platform and a shared corpus of multi-feature annotated music which will be reusable by both partners in the future and eventually disseminated

Accurate Real-time Windowed Time Warping
Robert Macrae and Simon Dixon

ABSTRACT – Dynamic Time Warping (DTW) is used to find alignments between two related streams of information and can be used to link data, recognise patterns or find similarities. Typically, DTW requires the complete series of both input streams in advance and has quadratic time and space requirements. As such DTW is unsuitable for real-time applications and is inefficient for aligning long sequences. We present Windowed Time Warping (WTW), a variation on DTW that, by dividing the path into a series of DTW windows and making use of path cost estimation, achieves alignments with an accuracy and efficiency superior to other leading modifications and with the capability of synchronising in real-time. We demonstrate this method in a score following application. Evaluation of the WTW score following system found 97.0% of audio note onsets were correctly aligned within 2000 ms of the known time. Results also show reductions in execution times over state-of-the- art efficient DTW modifications.

Idea: Frame window features – (sub dtw frames).  Each path can be calculated sequentially, so less history needs to be retained which is important for performance.

Works in linear time like previous systems, but with the smaller history it can work entirely in memory, so it avoids the problem of needing to store the history on disk. Nice demo of a real-time time warping.

Understanding Features and Distance Functions for Music Sequence Alignment – Ozgur Izmirli and Roger Dannenberg

ABSTRACT We investigate the problem of matching symbolic representations directly to audio based representations for applications that use data from both domains. One such application is score alignment, which aligns a sequence of frames based on features such as chroma vectors and distance functions such as Euclidean distance. Good representations are critical, yet current systems use ad hoc constructions such as the chromagram that have been shown to work quite well. We investigate ways to learn chromagram-like representations that optimize the classification of “matching” vs. “non-matching” frame pairs of audio and MIDI. New representations learned automatically from examples not only perform better than the chromagram representation but they also reveal interesting projection structures that differ distinctly from the traditional chromagram.

Roger and  Ozgur present a method for learning features for score alignment.  They bypass the traditional chromagram feature with a feature that is learned projection of the audio spectrum.  Results show that the new features work better than chroma.

Learning Tags that Vary Within a Song
Michael I. Mandel, Douglas Eck and Yoshua Bengio

Abstract:This paper examines the relationship between human generated tags describing different parts of the same song. These tags were collected using Amazon’s Mechanical Turk service. We find that the agreement between different people’s tags decreases as the distance between the parts of a song that they heard increases. To model these tags and these relationships, we describe a conditional restricted Boltzmann machine. Using this model to fill in tags that should probably be present given a context of other tags, we train automatic tag classifiers (autotaggers) that outperform those trained on the original data.

Michael enlisted the Amazon Turk to tag music. They paid for about a penny per tag.  About 11% of the tags were spam.  He then looked at co-occurence data that showed interesting patterns, especially as related to the distance between clips in a single track.

Results shows that smoothing of the data with a boltzmann machine tends to give better accuracy when tag data is sparse.

Sparse Multi-label Linear Embedding Within Nonnegative Tensor Factorization Applied to Music Tagging Yannis Panagakis, Constantine Kotropoulos and Gonzalo R. Arce

Abstract: A novel framework for music tagging is proposed. First, each music recording is represented by bio-inspired auditory temporal modulations. Then, a multilinear subspace learning algorithm based on sparse label coding is developed to effectively harness the multi-label information for dimensionality reduction. The proposed algorithm is referred to as Sparse Multi-label Linear Embedding Non- negative Tensor Factorization, whose convergence to a stationary point is guaranteed. Finally, a recently proposed method is employed to propagate the multiple labels of training auditory temporal modulations to auditory temporal modulations extracted from a test music recording by means of the sparse l1 reconstruction coefficients. The overall framework, that is described here, outperforms both humans and state-of-the-art computer audition systems in the music tagging task, when applied to the CAL500 dataset.

This paper gets the ‘Title that rolls off the tongue best’ award.   I don’t understand all of the math for this one, but some notes – the wavelet-based features used, seem to be good at discriminating at the genre level.  He compares the system to Doug Turnbull’s MixHier and to the system that we built at Sun labs with Thierry, Doug, Francois and myself (Autotagger: A model for predicting social tags from acoustic features on Large Music Databases)

.