Speech Recognition (SR) is a process in which spoken words or sentences are converted to text. In this process an acoustic signal, captured by a microphone, is converted to text. The previous couple decades have seen such rapid developments in automatic speech recognition systems. They are being used in various industries such as healthcare, agriculture, speech dialling, directory services, education, and autos, among others, most commonly being virtual assistants like Siri, Google Assistant or Cortana. Even after so much advancements these ASR systems are not commonly built for Indian regional language. Since, majority of the population of India speaks Hindi it becomes important for us to have an ASR framework for Hindi to improve the situation of the service sector. This paper presents an innovative approach for ‘Automatic Speech Recognition’ that uses a modified DeepSpeech2 model adapted for the Hindi language. The model integrates “Recurrent Neural Networks” (RNNs) and “Convolutional Neural Networks” (CNNs) for feature extraction with Gated Recurrent Units (GRUs) for sequence modelling. To further enhance performance, an attention mechanism is incorporated, allowing the model to focus on audio sequence segments. The purpose of this paper is to improve speech recognition performance for the Hindi language by utilising a modified DeepSpeech2 model along with some fine-tuning. The proposed method effectively models the Hindi language and reduces the training and validation loss with the help of efficient pre-processing and fine-tuning of the model.