COMPARISON METRICS AND PERFORMANCE ESTIMATIONS FOR DEEP BEAMFORMING DEEP NEURAL NETWORK BASED AUTOMATIC SPEECH RECOGNITION SYSTEMS USING MICROPHONE-ARRAYS

Automatic Speech Recognition (ASR) functionality, the automatic translation of speech into text, is on the rise today and is required for various use-cases, scenarios, and applications. An ASR engine by itself faces difficulties when encountering live input of audio data, regardless of how sophisticated and advanced it may be. That is especially true, under the circumstances such as a noisy ambient environment, multiple speakers, or faulty microphones. These kinds of challenges characterize a realistic scenario for an ASR system. ASR functionality continues to evolve toward more comprehensive End-to-End (E2E) solutions. E2E solution development focuses on three significant characteristics. The solution has to be robust enough to show endurance against external interferences. Also, it has to maintain flexibility, so it can easily extend in expectation of adapting to new scenarios or in order to achieve better performance. Lastly, we expect the solution to be modular enough to fit into new applications conveniently. Such an E2E ASR solution may include several additional micro-modules of speech enhancements besides the ASR engine, which is very complicated by itself. Adding these micro-modules can enhance the robustness and improve the overall system’s performance. Examples of such possible micro-modules include noise cancellation and speech separation, multi-microphone arrays, and adaptive beamformer(s). Being a comprehensive solution built of numerous micro-modules is technologically challenging to implement and challenging to integrate into resource-limited mobile systems. By offloading the complex computations to a server on the cloud, the system can fit more easily in less capable computing devices. Nevertheless, that compute offloading comes with the cost of giving up on real-time analysis, and increasing the overall system bandwidth. In addition, offloading to a server must have connectivity to the cloud over the internet. To find the optimal trade-offs between performance, Hardware (HW) and Software (SW) requirements or limitations, maximal computation time allowed for real-time analysis, and the detection accuracy, one should first define the different metrics used for the evaluation of such an E2E ASR system. Secondly, one needs to determine the extent of correlation between those metrics, plus the ability to forecast the impact each variation has on the others. This research presents novel progress in optimally designing a robust E2E-ASR system targeted for mobile, resource-limited devices. First, we describe evaluation metrics for each domain of interest, spread over vast engineering subjects. Here, we emphasize any bindings between the metrics across domains and the degree of impact derived from a change in the system’s specifications or constraints. Second, we present the effectiveness of applying machine learning techniques that can generalize and provide results of improved overall performance and robustness. Third, we present an approach of substituting architectures, changing algorithms, and approximating complex computations by utilizing a custom dedicated hardware acceleration in order to replace the traditional state-of-the-art SW-based solutions, thus providing real-time analysis capabilities to resource-limited systems