Hybrid TDNN
Hybrid TDNNs are a prominent architecture in automatic speech recognition (ASR), aiming to improve accuracy and efficiency. Current research focuses on optimizing these models through techniques like neural architecture search to automatically determine optimal network configurations, mixed-precision quantization to reduce model size and computational cost without significant performance loss, and system combination strategies that leverage the complementary strengths of hybrid TDNNs and end-to-end models like Conformers. These advancements contribute to creating more robust, resource-efficient ASR systems with potential for broader deployment in practical applications.
Papers
Towards Green ASR: Lossless 4-bit Quantization of a Hybrid TDNN System on the 300-hr Switchboard Corpus
Junhao Xu, Shoukang Hu, Xunying Liu, Helen Meng
Two-pass Decoding and Cross-adaptation Based System Combination of End-to-end Conformer and Hybrid TDNN ASR Systems
Mingyu Cui, Jiajun Deng, Shoukang Hu, Xurong Xie, Tianzi Wang, Shujie Hu, Mengzhe Geng, Boyang Xue, Xunying Liu, Helen Meng