In recent years, studies on automatic speech recognition (ASR) have shown outstanding results that reach human parity on short speech segments. However, there are still difficulties in standardizing the output of ASR such as capitalization and punctuation restoration for long-speech transcription. The problems obstruct readers to understand the ASR output semantically and also cause difficulties for natural language processing models such as NER, POS and semantic parsing. In this paper, we propose a method to restore the punctuation and capitalization for long-speech ASR transcription. The method is based on Transformer models and chunk merging that allows us to (1), build a single model that performs punctuation and capitalization in one go, and (2), perform decoding in parallel while improving the prediction accuracy. Experiments on British National Corpus showed that the proposed approach outperforms existing methods in both accuracy and decoding speed.

No abstract

Automatic Speech Recognition (ASR) systems convert human speech into the corresponding transcription automatically. They have a wide range of applications such as controlling robots, call center analytics, voice chatbot. Recent studies on ASR for English have achieved the performance that surpasses human ability. The systems were trained on a large amount of training data and performed well under many environments. With regards to Vietnamese, there have been many studies on improving the performance of existing ASR systems, however, many of them are conducted on a small-scaled data, which does not reflect realistic scenarios. Although the corpora used to train the system were carefully design to maintain phonetic balance properties, efforts in collecting them at a large-scale are still limited. Specifically, only a certain accent of Vietnam was evaluated in existing works. In this paper, we first describe our efforts in collecting a large data set that covers all 3 major accents of Vietnam located in the Northern, Center, and Southern regions. Then, we detail our ASR system development procedure utilizing the collected data set and evaluating different model architectures to find the best structure for Vietnamese. In the VLSP 2018 challenge, our system achieved the best performance with 6.5% WER and on our internal test set with more than 10 hours of speech collected real environments, the system also performs well with 11% WER

No abstract

No abstract