The TNT Team System Descriptions of Cantonese, Mongolian and Kazakh for IARPA OpenASR21 Challenge

This paper presents our systems and experimental analyses for the OpenASR21 Challenge. We describe the systems in the constrained condition, constrained-plus condition, and unconstrained condition, and our post-evaluation analyses for the Challenge. The systems in constrained condition are built based on hybrid automatic speech recognition (ASR) technology. For constrained-plus condition, pre-training models and system fusions are incorporated. We pre-train our conformer encoders using the wav2vec 2.0 pre-training method, which is called Ch-w2v-conformer. For unconstrained condition, our end-to-end ASR systems with conformer in optimized loss and long sequence encoders are adopted. We also adapt these challenging public switched telephone network (PSTN) conditions using publicly available data in the shape of wideband dictated speech with a similar accent, respectively. Finally, a series of systems are submitted for this challenge. Our submitted systems have achieved excellent performances for the three conditions.