Enhance run-time performance with a collaborative distributed speech recognition framework

This paper presents an improvement of a distributed Thai speech recognizer, aiming to enhance system response time as measured by a real-time factor (RTF) for a better user experience. The system is designed based on a collaborative multi-agents and task workers concept. A Streaming Agent is introduced to manage speech signal transfer while a Recognition Agent is applied to manage speech recognition task distribution. The speech recognition task is distributed to an available pipeline of task workers, which contain speech recognition core engines. A concept of task worker is introduced to provide light-weight management for each individual task in the pipeline. Both multi-agents and task workers are designed to work synchronously in order to minimize the overall processing time, especially in narrow-band or unstable network environment. The proposed improved system is compared with a traditional system in terms of their recognition word error rate (WER) and RTF. The results show that the implementation of speech codec, multi-agents and task workers in the proposed framework can substantially reduce the computational cost in terms of RTF by 42.7% on average in a narrow-band mobile network. In addition, there is no significant difference in WER between the proposed and baseline systems.