Increasing the Accessibility of Time-Aligned Speech Corpora with Spokes Mix

Increasing the Accessibility of Time-Aligned Speech Corpora with Spokes Mix Piotr P˛ezik University of Lodz Pomorska 171/173, 90-236 Lodz, Poland piotr.pezik@uni.lodz.pl Abstract Spokes Mix is an online service providing access to a number of spoken corpora of Polish, including three newly released time-aligned collections of manually transcribed spoken-conversational data. The purpose of this service is two-fold. Firstly, it functions as a programmatic interface to a number of unique collections of conversational Polish and potentially also spoken corpora of other languages, exposing their full content with complete metadata and annotations. Equally important, however, is its second function of increasing the general accessibility of these resources for research on spoken and conversational language by providing a centralized, easy-to-use corpus query engine with a responsive web-based user interface. Keywords: Speech corpora, corpus search engine, online language services 1. Introduction High-quality corpora of spoken conversational language are expensive to acquire. Worse still, even when they are col- lected, annotated and made publicly available, their poten- tial as sources of primary data may remain largely unreal- ized, due to the lack of search and exploration tools which could deal with the identification of speech-specific linguis- tic and multimodal phenomena. Apart from their research value, spoken language resources are also indispensable in the development of speech recognition systems and related language technologies. However, in order to sufficiently serve this purpose, they have to be standardized and in- teroperable enough to be used for both online and offline processing. Spokes Mix is an online service developed as part of the CLARIN-PL infrastructure to address both of these challenges for Polish. Firstly, the search engine of Spokes Mix makes it possible to explore not only the lin- guistically annotated transcriptions of spoken language, but also some characteristics of the underlying speech signal, such as duration and prosodic contours. These function- alities can be accessed through a web application, which also makes it relatively easy to customize, export and save search results by non-technical users. Secondly, the entire contents of the indexed corpora is available for program- matic search clients and it can exported on demand together with the time-aligned sound files. The present paper briefly describes both the newly and previously available corpora available through Spokes Mix. Next, some of the features of its search engine and API are described to illustrate its accessibility as a language resource service for both tech- nical and casual users. Finally, some of the key planned improvements of Spokes Mix are also briefly outlined. 2. The Data Spokes Mix provides access to a number of spoken corpora, including three entirely new collections of conversational Polish. To demonstrate and test the ability of Spokes Mix to simultaneously serve speech corpora in multiple languages, it also includes a collection of well-known speech corpora of English. All of these datasets are briefly specified below. 2.1. Annotation and Interoperability The corpora currently indexed in Spokes Mix have been unified with respect to the varying degrees of bibliographic metadata and sociolinguistic annotation they originally contained. Every transcription …

Content truncated to 500 words