Improving Amharic Speech Recognition System Using Connectionist Temporal Classification with Attention Model and Phoneme-Based Byte-Pair-Encodings

Out-of-vocabulary (OOV) words are the most challenging problem in automatic speech recognition (ASR), especially for morphologically rich languages. Most end-to-end systems performed at word and character levels of a language. Amharic is poorly resourced but This paper proposes hybrid connectionist temporal classification with attention architecture syllabification algorithm system (AASR) using its phoneme-based subword units. helps to insert epithetic vowel እ[ɨ], which not included our Grapheme-to-Phoneme (G2P) conversion developed consonant–vowel (CV) representations graphemes. The proposed model was trained various subwords, namely characters, phonemes, character-based subwords generated by byte-pair-encoding (BPE) segmentation algorithm. Experimental results showed that context-dependent tend result more accurate than character-based, phoneme-based, counterparts. Further improvement also obtained SpecAugment data augmentation technique. error rate (WER) reduction 18.38% compared acoustic modeling word-based recurrent neural network language (RNNLM) baseline. These models useful improve machine translation tasks.