Multi-Lingual Contextual Audio Transcription and Translation to Selected Native Languages

Abstract

This paper presents a contextual audio transcription and multilingual translation system designed for native languages. Utilizing advanced technologies such as AssemblyAI for accurate audio-to-text conversion, BERTopic for contextual topic modeling, and OpenAI’s API for indigenous language translation, the system demonstrates exceptional performance. AssemblyAI achieves low word error rates (WER), outperforming other transcription models. BERTopic effectively extracts meaningful topics, surpassing traditional models like Latent Dirichlet Allocation (LDA) in coherence and interpretability. The translation component, powered by GPT-4, produces accurate and contextually appropriate translations with low perplexity scores. This integrated approach bridges communication gaps in multilingual and multicultural contexts, offering a valuable tool for educators, professionals, and content creators to promote inclusivity and digital accessibility. While challenges with noisy data and language diversity remain, this system highlights the effectiveness of combining transcription, topic modeling, and translation tasks.