Development of High-Performance and Large-Scale Vietnamese Automatic Speech Recognition Systems

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

Automatic Speech Recognition of Vietnamese for a New Large-Scale Corpus

“…Table 1. The differences among tones in Vietnamese are illustrated in Table 2 [7]. The monosyllabic nature of Vietnamese, combined with its tonal system, adds a layer of complexity to the language.…”

“…All audio files were converted to the wave format with a sampling frequency of 16 kHz and PCM 16 bits. In [7], three Vietnamese speech corpora have been introduced. Those corpora include two small reading speech corpora with a total of 6 h and 6.5 h, respectively, and a large-scale speech corpus with 900 h. The large-scale speech corpus was collected by crawling untranscripted audio from various resources, such as movies, YouTube movies, and electronic newspapers.…”

“…In the past twenty years, there have been a lot of efforts to increase the number of Vietnamese speech corpora, but the number of large-scale Vietnamese speech corpora has been limited. Most public corpora of Vietnamese speech are reading speech, like the FTRI corpus and the corpora in [2,3,8], or part of reading speech, such as VinBigdata-VLSP2020 or two small corpora in [7]. Therefore, the available corpora might not be compatible with real-life scenarios for spoken language, like conversational and discussion recognition.…”

“…Many current Vietnamese corpora have a small size, around a few hours to tens of hours, such as VIVOS, VLSP 2018, etc., [2,3,5]. The Vietnamese corpora of more than 100 h, such as VinBigdata-VLSP2020 and corpora in [6,7], are rare, and most of them are not open-access, like the corpus collected by FPT Technology Research Institute (FTRI), namely FTRI corpus, MICA VNSpeechCorpus [8], and the corpora in [6,7]. However, those corpora are not either high-quality sound or open-access.…”

“…Most previous works for Vietnamese automatic speech recognition (ASR) tasks investigated traditional statistical ASR models like HMM/GMM, deep neural networks (DNNs) or hybrid models-a combination of HMM/GMM and DNN models [3,6,7,[9][10][11]. Recently, end-to-end (E2E) models in the field of automatic speech recognition have gained considerable attention from both academic and industrial perspectives [12][13][14][15][16][17].…”