MFLA: Monotonic Finite Look-ahead Attention for Streaming Speech Recognition
This addresses the problem of enabling streaming applications for speech recognition systems, representing an incremental improvement by adapting existing models with novel mechanisms.
The paper tackles the challenge of integrating large pre-trained speech models like Whisper into streaming systems by proposing a prefix-to-prefix training framework with a Continuous Integrate-and-Fire mechanism and Monotonic Finite Look-ahead Attention, achieving a controllable trade-off between latency and quality for streaming speech recognition.
Applying large pre-trained speech models like Whisper has shown promise in reducing training costs for various speech tasks. However, integrating these models into streaming systems remains a challenge. This paper presents a novel prefix-to-prefix training framework for streaming recognition by fine-tuning the Whisper. We introduce the Continuous Integrate-and-Fire mechanism to establish a quasi-monotonic alignment between continuous speech sequences and discrete text tokens. Additionally, we design Monotonic Finite Look-ahead Attention, allowing each token to attend to infinite left-context and finite right-context from the speech sequences. We also employ the wait-k decoding strategy to simplify the decoding process while ensuring consistency between training and testing. Our theoretical analysis and experiments demonstrate that this approach achieves a controllable trade-off between latency and quality, making it suitable for various streaming applications.