CLOUD-BASED DEEP NEURAL NETWORKS COMBINED WITH LOCAL SPEECH RECOGNITION DECODER

摘要

Cloud-based deep neural networks combined with local decoder for automatic speech recognition belong to the field of digital speech signal processing, specifically to automatic speech recognition. Existing systems do not combine server and local possibilities and thus are either not specific enough for a given task or lack resources for supporting structures which are sufficiently complex to achieve a satisfcatory level of recognition accuracy and speed. This invention enables speech recognition with good performance within limited domains of interaction by using a combination of powerful computational resources at the server side and flexible local resources at the client side. Computational resources of the server are used to support functioning of neural networks of appropriate complexity, i.e. acoustic speech modelling, while, on the other side, specific lexicons and language models are used to personalize the speech recognition system at each user device. Client devices are used to perform only decoding, as the last step in speech recognition. The described system enables reliable speech recognition within a given domain, even on computationally less powerful platforms and in less than ideal ambiental conditions.