natural-language-process

Arxiv — summary generated by Brevi AssistantIt’s challenging to personalize transducer-based automated speech recognition system with context information which is inaccessible and vibrant during version training. Experiments reveal that the design improves standard ASR model performance with about 50% relative word mistake rate decrease, which also substantially outperforms the baseline approach such as contextual LM biasing. In this paper, we provide AISHELL-4, a sizable real-recorded Mandarin speech dataset gathered by 8-channel round microphone selection for speech processing in conference scenario. Provided most open resource dataset for multi-speaker tasks are in English, AISHELL-4 is the only Mandarin dataset for conversation speech, supplying added worth for information diversity in...