Bringing Order to Chaos: A Non-Sequential Approach for Browsing Large Sets of Found Audio Data

Bringing Order to Chaos: A Non-Sequential Approach for Browsing Large Sets of Found Audio Data Per Fallgren, Zofia Malisz, Jens Edlund KTH Royal Institute of Technology perfall@kth.se, malisz@kth.se, edlund@speech.kth.se Abstract We present a novel and general approach for fast and efficient non-sequential browsing of sound in large archives that we know little or nothing about, e.g. so called found data – data not recorded with the specific purpose to be analysed or used as training data. Our main motivation is to address some of the problems speech and speech technology researchers see when they try to capitalise on the huge quantities of speech data that reside in public archives. Our method is a combination of audio browsing through massively multi-object sound environments and a well-known unsupervised dimensionality reduction algorithm (SOM). We test the process chain on four data sets of different nature (speech, speech and music, farm animals, and farm animals mixed with farm sounds). The methods are shown to combine well, resulting in rapid and readily interpretable observations. Finally, our initial results are demonstrated in prototype software which is freely available. Keywords: found data, data visualisation, speech archives 1. Introduction 1.1. Found data for speech technology The availability of usable data becomes ever more impor- tant as data-driven methods continue to dominate virtually every field. Numerous organisations (e.g. the Wikime- dia Foundation1, the World Wide Web Consortium2, and the Open Data Institute3) push hard for Open Data. Al- though language data, and speech data in particular, is rid- dled with complex legally restricting considerations (Ed- lund and Gustafson, 2016) and less likely to be ”non- privacy-restricted” and ”non-confidential” as required of Open Data, the use of data-driven methods in language technology (LT) and speech technology (ST) is nothing less than a modern success story. In the intersection of LT and other fields, such as history and politics, social sciences and health (Gregory and Ell, 2007; Sylwester and Purver, 2015; Zhao et al., 2016; Pestian et al., 2017), traditional data- driven methods play a significant role. Data is arguably yet more crucial in ST, and for decades, funding agencies have spent considerable resources on projects that record speech data. These efforts have been dwarfed by the vast amounts of user data that are being gathered by multinational corpo- rate giants for the betterment of their proprietary technolo- gies. In contrast, comparable amounts of data are not available to academia and smaller companies. As a result, at least when it comes to the LT and ST tasks that are targeted by the major commercial players, systems developed by smaller entities do not have a chance to compete. This resource gap raises concerns: what happens if the giants decide to charge large sums for their solutions once we have grown accustomed to getting them cheap? How does one conduct research that requires solutions to work on tasks different from those targeted by the giants? And how do we analyse data recorded under entirely different circumstances? As it 1https://en.wikipedia.org/wiki/Wikidata 2https://www.w3.org/ 3https://theodi.org/ stands, the …

Content truncated to 500 words