An Efficient Method for Improving Automatic Speech Recognition NUSRAT JAHAN 1 , MD. ASHIKUR RAHMAN KHAN 1, *, ZAYED US SALEHIN 1 , NISHU NATH 1 1 Department of Information and Communication Engineering Noakhali Science and Technology University Noakhali-3814 BANGLADESH Abstract: - Automatic speech recognition translates spoken words into the text; It is still a challenging task due to the high viability in speech signals. Several decoding algorithms and recognition systems have been developed, aimed at various recognition tasks. The design of the speech recognition system requires careful attention to the challenges or issue such as various types of speech classes, speech representation, feature extraction techniques, database and performance evaluation. This paper presents a study of basic approaches to speech recognition and also presents an error analysis of existing speech recognition system to provide a better system. .Keywords: Hidden Markov model, Acoustic model, language model, Feature Extraction, Google Web Speech API, Voice Notepad. Received: June 15, 2021. Revised: April 16, 2022. Accepted: May 17, 2022. Published: June 9, 2022. 1 Introduction Speakers may have different accents, dialects or pronunciations, and speak in different styles, at different rates, and in different emotional states. . For more than half a century, research has been conducted in the field of automatic speech recognition (ASR), which constitutes an important part in the fulfilment of this vision. Despite the considerable amount of research resources invested in this task, many questions remain to be answered. This is because the problem is very complex and requires solutions from several disciplines. Early attempts at ASR were based on template matching techniques. HMMs are suitable for acoustic modelling in a context of concatenated acoustic- phonetic units. Although there are deficiencies associated with these devices for acoustic modelling, they have proven effective for the processing of continuous speech. The trend in ASR has been toward increasingly complex models, to improve recognition accuracy and involve larger vocabularies. This paper focuses on error analysis of speech recognition systems. The trend in ASR has been toward increasingly complex models, to improve recognition accuracy and involve larger vocabularies. To handle these more complex recognition tasks, several advanced decoding strategies are required. The current challenges of speech recognition are caused by two major factors- reach and loud environments. The current challenges of speech recognitions are diverse. Speech recognition software isn't always able to interpret spoken words correctly. This is due to computers not being on par with humans in understanding the contextual relation of words and sentences, causing misinterpretations of what the speaker meant to say or achieve. People usually assume that computerizing a process would speed it up. Unfortunately, this is not always the case when it comes to voice recognition systems. In many cases using a voice, the app takes up more time than going with a traditional text-based version. While systems are getting better there’s still a big difference in their ability to understand American or Scottish English for example. Even a simple cold can be a reason for voice commands not to work as well as usual. keeping user data safe can easily become a conflict of interests. Therefore, a great challenge of voice recognition lies in making data input available for AI, but still, acknowledge the need for data privacy and security. Some things have to take into consideration. These are 1. Variability in speech 2. Vocal Range (pitch and format Frequency) 3. Age, Gender of the speaker 4. Voice Quality 5. Emotional State 6. Speech Style The main objective of this research is to study the accuracy, readability, accessibility of the speech recognition system. Also works with the problems or Nusrat Jahan et al. International Journal of Computers http://www.iaras.org/iaras/journals/ijc ISSN: 2367-8895 19 Volume 7, 2022