A Semi-Automatic Approach to Create Large Gender- and Age-Balanced Speaker Corpora: Usefulness of Speaker Diarization & Identification.

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3271–3280 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3271 A Semi-Automatic Approach to Create Large Gender- and Age-Balanced Speaker Corpora: Usefulness of Speaker Diarization & Identification. R´emi Uro1,2, David Doukhan1, Albert Rilliard2,3, La¨etitia Larcher1, Anissa-Claire Adgharouamane1, Marie Tahon4, Antoine Laurent4 1French National Institute of Audiovisual (INA), Paris, France. 2Universit´e Paris Saclay, CNRS, LISN. 3Universidade Federal do Rio de Janeiro, Brasil. 4LIUM, Le Mans Universit´e, France. {ruro,ddoukhan,llarcher,aadgharouamane}@ina.fr albert.rilliard@lisn.fr, {marie.tahon,antoine.laurent}@univ-lemans.fr Abstract This paper presents a semi-automatic approach to create a diachronic corpus of voices balanced for speaker’s age, gender, and recording period, according to 32 categories (2 genders, 4 age ranges and 4 recording periods). Corpora were selected at French National Institute of Audiovisual (INA) to obtain at least 30 speakers per category (a total of 960 speakers; only 874 have be found yet). For each speaker, speech excerpts were extracted from audiovisual documents using an automatic pipeline consisting of speech detection, background music and overlapped speech removal and speaker diarization, used to present clean speaker segments to human annotators identifying target speakers. This pipeline proved highly effective, cutting down manual processing by a factor of ten. Evaluation of the quality of the automatic processing and of the final output is provided. It shows the automatic processing compare to up-to-date process, and that the output provides high quality speech for most of the selected excerpts. This method shows promise for creating large corpora of known target speakers. Keywords: semi-automatic processing, corpus creation, diarization, speaker identification, gender-balanced, age-balanced, speaker corpus, diachrony 1. Introduction This paper describes a semi-automatic method for speeding-up speaker corpora construction. There is a growing need for reference speech corpora featur- ing reliable and known speaker characteristics. Man- ual annotations of high level features such as speaker segmentation is a time consuming process (Broux et al., 2018). We propose to use up-to-date diarization and speaker identification to reduce human interven- tion, and enable the creation of large spoken corpora at a minimum cost. This work is part of the Gender Equality Monitor (GEM) project, that aims at describing male and fe- male representation differences in the French broadcast media at scale, using automatic information extraction methods. A main institution involved in the project is the French National Institute of Audiovisual (INA), a public institution in charge of archiving French audio- visual heritage (so called legal deposit). INA’s collec- tions include 23 million hours of TV and radio pro- grams broadcasted since the 1930’s. First, the targeted characteristics of the corpus are pre- sented, with a discussion of comparable datasets, and the potential use of this resource. A literature review on diarization and speaker identification systems is then presented. The rest of the paper presents: (1) the methodological steps for the creation of this corpus, (2) evaluations of these steps in terms of performance on available data with comparable characteristics, (3) a subjective evaluation of the quality of voice of the final …

Content truncated to 500 words