Nadee07/Speech-to-text

🌍 Transformer-Based Speech Recognition System for Low-Resource Languages # 🌍 Transformer-Based Speech Recognition System for Low-Resource Languages ### 🎓 Mini Research Project – 2025 **Developed using TensorFlow, Keras, Flask, HTML, CSS, and JavaScript** --- ## 🧠 Overview Modern speech recognition models such as **Whisper** and **Wav2Vec2.0** have achieved great success on high-resource languages like English and Mandarin. However, these models often fail to perform well for **low-resource languages** due to limited training data, lack of linguistic diversity, and high computational costs. Our research addresses the **low-resource language recognition gap** by developing a **Transformer-based Speech Recognition System** optimized for **low-resource languages** using publicly available multilingual datasets and data augmentation techniques. --- ## 🚩 Problem Statement Current multilingual ASR (Automatic Speech Recognition) models are limited to around **100 languages**, whereas there are **over 7000 languages spoken worldwide**. Low-resource languages face significant challenges such as: - 🧩 **Data Scarcity** – A lack of high-quality, labeled audio-text pairs. - ⚙️ **Domain Mismatch** – Existing datasets often don't match the real-world use cases for these languages. - 💻 **Computational Constraints** – Large-scale ASR models demand significant GPU/TPU resources. - 🗣️ **Dialect Diversity** – A single language can have numerous dialects, further fragmenting available data. --- ## 💡 Research Gap | Identified Gap | Description | |----------------|-------------| | Limited Coverage | High-quality ASR systems trained on only a small subset of global languages. | | Lack of Balanced Datasets | Imbalanced multilingual corpora with dominance of English and European languages. | | Computational Barriers | Training large models on low-resource datasets is often infeasible. | | Dialect Variation | Absence of datasets that represent multiple dialects per language. | --- ## 🧩 Proposed Solution Our research introduces a **transformer-based A …