Automatic Speech Recognition

Our AI team focuses on cutting-edge artificial intelligence research and applications.

Introduction​

Automatic Speech Recognition (ASR), also known as Speech-to-Text (STT), is the task of transforming speech signals into written content. While not a new field, ASR remains a challenging area of research due to the complexities of natural language and acoustic variability.

Our research group investigates the theoretical and practical limits of ASR across two distinct computational paradigms:

1. Real-Time Spontaneous Speech Understanding (Low-Latency ASR) This research stream focuses on the "latency-accuracy trade-off." We aim to construct lightweight, highly optimized architectures capable of processing streaming audio with minimal delay.

Research Challenges: Robustness in spontaneous environments and contextual biasing —the ability to recognize domain-specific entities unseen during training without retraining the core model.

2. High-Fidelity Asynchronous Transcription (Large-Scale ASR) This stream prioritizes transcription precision and semantic coherence over immediate latency. We explore heavy-weight architectures designed for offline inference on long-form audio.

Research Challenges: Efficient processing of long-context audio sequences and developing unified systems capable of handling code-switching (e.g., Vietnamese-English mixing) and generalizing across diverse linguistic syntaxes.

Contact: Nguyen Thi Tra My | ✉️ my.hust225049@gmail.com

Research Direction​

• Contextualized ASR: Developing models that can precisely recognize context-sensitive and often "biased" terms, such as proper names, addresses, locations, and phone numbers, which are critical for accurate information extraction.
• Multilingual ASR: Creating models capable of recognizing speech content containing two or more languages within the same stream (code-switching), such as the Vietnamese-English mix often found in casual conversation (e.g., "enjoy cái moment này").
• New Architectures for ASR: Investigating novel architectural approaches, including self-conditioned models, the use of discrete tokens, and weakly supervised loss functions to improve model efficiency and generalization.
• LLM in ASR: Integrating Large Language Models (LLMs) into the ASR pipeline to leverage their semantic understanding for significantly improved transcription quality and correction.

Members​

Nguyen Thi Tra My

Team Leader

Vu Nhat Minh

Researcher