SBIR-STTR Award

This proposal explores feature representations for automatic speech processing algorithms with a focus on preserving the clients' privacy. We address two different use cases of privacy: a system that transcribes speech, but is unable to infer the identity of speakers, such as an anonymous tip hotline; and a system that identifies speakers, but is unable to recognize the communicated messages, useful in scenario where a many conversations may be under surveillance in order to locate and isolate a targeted individual. We investigate the feasibility of implementing such systems by focusing on the audio signal representations in a distributed system, where embedded devices compute representations and transmit them to a Big Data analytics service via the Internet. Basic spectral acoustic features, tandem/bottleneck features, and high-dimensional outputs from deep neural networks will all be evaluated for both automatic speech recognition (ASR) and speaker identification (SID) tasks. Performance will be assessed to identify configurations favorable for both use cases. We additionally consider the possibility that an adversarial service operator may attempt to associate identities of speakers by clustering received requests, or reconstructing messages that have been transmitted over a mixture network. In addition to serving the homeland security mission, the private-sector commercialization potential of this research is substantial. It could prove useful to Remeeting, which is being developed as a mobile app and cloud service to record personal conversations; significant privacy concerns serve as barriers to adoption by individual and enterprise customers.

The purpose of this Phase II SBIR project is to improve technical capabilities and develop commercial products that record spoken conversations in a privacy-protecting manner. The primary aim is to support the homeland security mission with solutions that enable more effective law enforcement while simultaneously protecting the privacy and civil liberties of individuals. A secondary aim is to pursue commercialization in the private sector, specifically in the context of Remeeting: a voice recorder that enables searchable conversations, to improve the effectiveness of recalling and sharing information from interviews and meetings. The proposed work builds upon Phase I results exploring audio signal representations that retain only speaker identity or message content, such that enabling automatic speaker identification or speech recognition is mutually exclusive. This capability is useful, for example, because it can minimize an individual's exposure to potential loss of privacy by protecting either their identity or the communicated message. Phase II will further apply automatic keyword spotting and automatic speaker identification to diverse use cases within the homeland security mission, such as audio from body-worn cameras and facilitating compliance with statutory requirements for electronic surveillance. Such capabilities could be used to trigger audio recording only upon detecting a configurable set of relevant words or speakers, with adjustable sensitivity controls. Additional applications include smart meeting room microphones that can be configured to record in certain contexts, as well as ad-hoc microphone arrays using nearby mobile devices.