Microsoft Technology Licensing, LLC

Salman Quazi

Pradip Kumar Fatehpuria

Travis Wilson

John Vicondoa

Christopher Hakan Basoglu

Sayan Dev Pathak

Girish Milind Mahajan

Harini Kesavamoorthy

Osama Abuelsorour

Valeriy Viktorovich Kirshin

Pradip Fatehpuria

A technique partitions a user's original query into plural smaller component queries, each of which has a common part and an instance-specific part. The technique distributes the component queries to plural processor instances of a processor. The plural processor instances transform the respective component queries into query-component responses by acting in parallel, independent of each other. The technique generates a final response based on the query-component responses, e.g., by assembling the component-query responses into the final response. The technique reduces latency because the processor instances work on parts of the user's original query at the same time, rather than as a single stream of consecutive tokens. The plural processor instances have access to a shared cache memory, and utilize relevant data that has been computed in response to previous queries.

Reducing latency by processing parts of a language model query in parallel

Incorporation of an exogenous large-vocabulary model into rule-based speech recognition is provided. An audio stream is received by a local small-vocabulary rule-based speech recognition system (SVSRS), and is streamed to a large-vocabulary statistically-modeled speech recognition system (LVSRS). The SVSRS and LVSRS perform recognitions of the audio. If a portion of the audio is not recognized by the SVSRS, a rule is triggered that inserts a mark-up in the recognition result. The recognition result is sent to the LVSRS. If a mark-up is detected, recognition of a specified portion of the audio is performed. The LVSRS result is unified with the SVSRS result and sent as a hybrid response back to the SVSRS. If the hybrid-recognition rule is not triggered, an arbitration algorithm is evoked to determine whether the SVSRS or the LVSRS recognition has a lesser word error rate. The determined recognition is sent as a response to the SVSRS.

Incorporating an exogenous large-vocabulary model into rule-based speech recognition

Growing community of inventors

San Jose, WA, United States of America

Salman Quazi