Analysis of multilingual sequence-to-sequence speech recognition systems

This paper investigates the applications of various multilingual approaches developed in conventional deep neural network - hidden Markov model (DNN-HMM) systems to sequence-to-sequence (seq2seq) automatic speech recognition (ASR). We employ a joint connectionist temporal classification-attention network as our base model. Our main contribution is separated into two parts. First, we investigate the effectiveness of the seq2seq model with stacked multilingual bottle-neck features obtained from a conventional DNN-HMM system on the Babel multilingual speech corpus. Second, we investigate the effectiveness of transfer learning from a pre-trained multilingual seq2seq model with and without the target language included in the original multilingual training data. In this experiment, we also explore various architectures and training strategies of the multilingual seq2seq model by making use of knowledge obtained in the DNN-HMM based transfer-learning. Although both approaches significantly improved the performance from a monolingual seq2seq baseline, interestingly, we found the multilingual bottle-neck features to be superior to multilingual models with transfer learning. This finding suggests that we can efficiently combine the benefits of the DNN-HMM system with the seq2seq system through multilingual bottle-neck feature techniques.