Transcription System for Regional Accents in Angola Using Whisper and LLM

Over the years, speech recognition technology has evolved significantly. Many artificial intelligence models that have been developed do not take into account the nuances of a given language nor the context of the conversation. Specifically for the Portuguese language, the dataset used in the training process of several models is relatively small compared to other languages such as English, Spanish, etc. In this paper, we propose an automatic speech transcription solution for Portuguese spoken in Angola, which considers the different accent variations and the conversational context. For the development of this solution, we performed fine-tuning of the Whisper speech recognition model and used a large language model for transcription correction and context identification. The fine-tuning was carried out in Google Colab mainly due to its accessibility.