The Voice User Interface (VUI) for human-computer interaction has received wide acceptance, due to which the systems for speech recognition in regional languages are now being developed, taking into account all of the dialects. Because of the limited availability of the speech corpus (SC) of regional languages for doing research, designing a speech recognition system is challenging. This contribution provides a Parallel Big Bang-Big Crunch (PB3C)-based mechanism to automatically evolve the optimal architecture of LSTM (Long Short-Term Memory). To decide the optimal architecture, we evolved a number of neurons and hidden layers of LSTM model. We validated the proposed approach on Marathi speech recognition system. In this research work, the performance comparisons of the proposed method are done with BBBC based LSTM and manually configured LSTM. The results indicate that the proposed approach is better than two other approaches.