Buffered Whisper Transformer: A Real-time Speech Recognition System for Edge Devices via Dynamic Audio Chunking Mechanism

By Fei PAN, Yechuan ZHOU, Junhao ZHANG, Yuchen TAN

Highlights

• Empowers non-streaming recognition models with accurate real-time speech recognition capabilities

• Dynamic audio chunking with buffering queue enabling uninterrupted processing and continuous result updates

• Effectively mitigates semantic fragmentation common in commercial streaming recognition systems

• Robust multilingual and dialect recognition

Applications

• Unlocks broad applications demanding high-precision recognition, such as simultaneous interpretation and live captioning

• Enables smart home and in-vehicle voice system solutions with improved responsiveness and reliability