Text-to-Speech Conversation Using Optimized Deep Learning Model

This research uses text-to-speech (TTS) pretraining to provide a unique transformer-based sequence-to-sequence voice conversion (VC) model. The capacity of seq2seq VC models to translate prosody is a major selling point. Although VC has been successfully applied to seq2seq models based on RNNs and CNNs, the usage of the transformer system, which has shown promising results in many speech processing applications, has not yet been explored. However, seq2seq models are impractical because of their need for a large amount of data and because the converted speech is often mispronounced. To address these issues, we propose a refined (LSTM)-based time-series method in this work. An enhanced LSTM model is central to the proposed system, since it has the potential to improve time-series prediction across a range of devices. The hyper-parameters of the enhanced LSTM are best selected using an enhanced sparrow search technique. The mutation operative was utilized to enhance the position update of the joiner and prevent the algorithm from descending into the local dangerous value, while the learning constant was inserted in the location update portion of the discoverer to enhance the algorithm's global search capabilities. When given such pre-trained model parameters, VC models are able to effectively construct hidden representations for natural-sounding, understandable transformed speech. The experimental results demonstrate that such a pretraining approach is superior to an LSTM-based ISSA VC model in terms of intelligibility, naturalness, and similarity, and that it can assist data-efficient training.