WeCanTalk: A New Multi-language, Multi-modal Resource for Speaker Recognition

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3451–3456 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3451 WeCanTalk: A New Multi-language, Multi-modal Resource for Speaker Recognition Karen Jones, Kevin Walker, Christopher Caruso, Jonathan Wright, Stephanie Strassel Linguistic Data Consortium University of Pennsylvania, Philadelphia, USA {karj, walkerk, carusocr, jdwright, strassel}@ldc.upenn.edu Abstract The WeCanTalk (WCT) Corpus is a new multi-language, multi-modal resource for speaker recognition. The corpus contains Cantonese, Mandarin and English telephony and video speech data from over 200 multilingual speakers located in Hong Kong. Each speaker contributed at least 10 telephone conversations of 8-10 minutes’ duration collected via a custom telephone platform based in Hong Kong. Speakers also provided at least 3 videos in which they were both speaking and visible, along with one selfie image. At least half of the calls and videos for each speaker were in Cantonese, while their remaining recordings featured one or more different languages. Both calls and videos were made in a variety of noise conditions. All speech and video recordings were manually audited for quality including presence of the expected language and for speaker identity. The WeCanTalk Corpus has been used to support the NIST 2021 Speaker Recognition Evaluation and will be published in the LDC catalog. Keywords: speaker recognition, speech database, video, telephony, Cantonese, Mandarin, Chinese, English 1. Introduction The WeCanTalk (WCT) Corpus is a multi-modal, multi- language corpus comprising conversational telephone speech (CTS) and audio from video (AfV) recordings from native Cantonese speakers recruited in Hong Kong who were also fluent in at least one other language. The corpus was designed to support the development and evaluation of speaker recognition technologies. The collection consists of data from 202 distinct speakers, with at least 10 telephone calls and 3 videos plus an accompanying selfie image provided by each speaker. All speakers made at least 5 of their 10 required telephone calls and at least 2 of their 3 required videos in Cantonese. The remaining non-Cantonese recordings, whether video or call, were required to be in a different language that the participant spoke fluently. In most cases the speaker’s secondary language was either English or Mandarin. Telephone calls were made and recorded in Hong Kong via a custom-built telephone collection platform designed by Linguistic Data Consortium (LDC) and built and operated by Hong Kong Polytechnic University (PolyU) under LDC’s direction. Video recordings and selfie images produced by each speaker were uploaded to a custom website developed and hosted by LDC. WCT data collection took place in June-October 2020, a time characterized by significant political and social turmoil in Hong Kong including widespread political protests and increasing impact of the COVID-19 pandemic, both which affected the collection approach, particularly for video data, as well as the overall project timeline. Despite these challenges, the collection was successfully completed and the resulting corpus was used in the NIST 2021 Speaker Recognition Evaluation (NIST, 2021; Sadjadi et al, 2022). The sections that follow describe the WCT …

Content truncated to 500 words