Bidirectional Long-short Term Memory Based Text to Speech Synthesis for Amharic Language

No Thumbnail Available



Journal Title

Journal ISSN

Volume Title


Addis Ababa University


Text-to-speech (TTS) synthesis is the automatic conversion of written text to spoken language. TTS systems show an imperative character in natural human-computer interaction. The aim of this work is to develop a bidirectional long-short term based TTS system for the Amharic Language. The system has two phases, the training and synthesis phases. In the training phase, first the text normalization is done and then from the normalized text linguistic features are extracted by using festival tool and the extracted features are used as input for the BLSTM based duration model. Then after that, duration model training is done and the model adds duration information on the extracted linguistic features and feeds for the BLSTM based acoustic model. The world vocoder extracts many acoustic frames composed of features which describe the signal in a more convenient way and used as an input for the acoustic model. Aco5ustic model training is done to map the input linguistic features and the associated duration features into acoustic features. We have prepared 600 speech their corresponding text transcription from Amharic audio bible by a male speaker. For this work the open source merlin speech synthesis toolkit, festival speech synthesis tool as a frontend and world vocoder are used. We have also prepared a pronunciation dictionary (lexicon) of 2500 words, phone set, letter to sound rule and question file set for frontend text processing based on the phonetic structure of Amharic language. In order to test the performance of the system we have performed subjective and objective evaluation. The evaluation with a listening test by 10 volunteers gave a score in MOS of 3.8 for intelligibility and 3.9 for naturalness to our BLSTM model and 3.65 for intelligibility and 3.7 for naturalness to our DNN model and MCD of BLSTM and DNN is 4.68 and 4.7 respectively.



Deep Learning, Recurrent Neural Networks, Long-Short Term Memory, Duration Model, Acoustic Model, Vocoder, Linguistic Features, Acoustic Features