Far-Field Speaker Recognition Benchmark Derived From The DiPCo Corpus

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1955–1959 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1955 Far-Field Speaker Recognition Benchmark Derived From The DiPCo Corpus Mickael Rouvier, Mohammad Mohammadamini LIA - Avignon University 339 Chemin des Meinajaries, 84000, Avignon {mickael.rouvier, mohammad.mohammadamini}@univ-avignon.fr Abstract In this paper, we present a far-field speaker recognition benchmark derived from the publicly-available DiPCo corpus. This corpus comprises three different tasks that involve enrollment and test conditions with single- and/or multi-channel recordings. The main goal of this corpus is to foster research in far-field and multi-channel text-independent speaker recognition. Also, it can be used for other tasks such as dereverberation, denoising, and speech enhancement. In addition, we release recipes implemented with Kaldi and SpeechBrain libraries to facilitate further research. We validate the evaluation design with a single-microphone state-of-the-art speaker recognition system (i.e. ResNet-101). The results show that the proposed tasks are very challenging. And we hope these resources will inspire the speech community to develop new methods and systems for this challenging domain. Keywords: speaker recognition, benchmark, far-field, DiPCo corpus 1. Introduction Speaker Recognition refers to the authentication of the claimed users from their voice (Bimbot et al., 2004). Speaker recognition systems have been used in several applications such as speaker diarization (Rouvier and Meignier, 2012), forensics (Campbell et al., 2009) or voice dubbing (Gresse et al., 2017). Although the state- of-the-art DNN-based systems perform well in adverse environments, still there are some challenges that re- duce their performance dramatically. Far-field speaker recognition is among the well-known challenges facing speaker recognition systems. The far-field challenge is intertwined with other challenges because the far dis- tance signal will be impacted more by other distortions such as noise and reverberation. Having a real-world evaluation benchmark for far-field speaker recognition is highly demanded by the speech community. Over the years, a few far-field corpus or far-field challenges have emerged. In (Qin et al., 2020), a Far-Field Speaker Verifica- tion Challenge (FFSVC) was proposed with a special focus on far-field distributed microphone arrays under noisy conditions in real scenes. In FFSVC three tasks are proposed: far-field text-dependent speaker verifi- cation from a single microphone array, far-field text- independent speaker verification from a single micro- phone array, and far-field text-dependent speaker veri- fication from distributed microphone arrays. The VOiCES challenge (Nandwana et al., 2019) fo- cuses on the robustness of the replayed speech in the presence of reverberation and background noises. VOiCES corpus was collected by recording played au- dio from high-quality loudspeakers in real rooms, cap- turing natural reverberation. The retransmitted cor- pus is LibriSpeech which is used as the clean speech source, while television, music, or babble played si- multaneously from another loudspeaker as background noise. The main deficiency of VOiCES is that it doesn’t match real-world scenarios. In (Garcia-Romero et al., 2019), the authors propose a speaker recognition benchmark derived from the CHIME-5 challenge. The benchmark comprises four tasks for enrollment and test conditions with single- speaker …

Content truncated to 500 words