Decoupling recognition and transcription in Mandarin ASR

Abstract

Much of the recent literature on automatic speech recognition (ASR) is taking an end-to-end approach. Unlike English where the writing system is closely related to sound, Chinese characters (Hanzi) represent meaning, not sound. We propose factoring audio → Hanzi into two sub-tasks: (1) audio → Pinyin and (2) Pinyin → Hanzi, where Pinyin is a system of phonetic transcription of standard Chinese. Factoring the audio → Hanzi task in this way achieves 3.9% CER (character error rate) on the Aishell-1 corpus, the best result reported on this dataset so far.

SCaLa: Supervised Contrastive Learning for End-to-End Speech Recognition

“…As shown in Table 1, we compare SCaLa with state-of-the-art ASR systems including hybrid [28], end-to-end [5,29], and selfsupervised learning [10,30]. Numerically, SCaLa outperforms the traditional CTC models [20] with 2.84% and 1.38% CER reductions on reading and spontaneous speech data, respectively.…”

“…4.3. Experimental results also show that SCaLa significantly outperforms hybrid chain models [28], end-to-end CTC-Conformer systems [29], self-supervised systems learning [10,30], and methods with phoneme masking [5].…”

“…Aishell-1 JD-tel Speaking style reading spontaneous Chain model [28] 7.45 15.94 Wav2vec2.0 [10,30] 5.30 16.30 WeNet CTC-conformer [29] w/ CTC prefix beam search 5.91 15.62 w/ attention rescoring 5.30 14.74 CTC [20] 6.74 15.34 CTC+phoneme mask [5] 5.11 14.80 SCaLa 3.90 13.96…”