System for recording and analyzing telephone conversations between humans and artificial intelligence

点击翻译按钮获取中文摘要

Assessing speech and language is an expensive and laborious process that is completely unsustainable given Canada’s rapidly changing demographics. In addition to direct costs of assessment, indirect costs include lost time in travel, wait times, and hours spent in assessment, which is a process so laborious that it is often repeated only every few years. This is the case across the population, from assessing children for autism, to assessing older adults for Alzheimer's disease (AD) and dementia. Repeatable, remote, and reliable assessment is essential. Many individuals have diminished language comprehension, vocabulary, and speech fluency. Our team has recently developed computational methods that are up to 100% accurate in identifying primary progressive aphasia, and over 81% accurate in identifying AD from spontaneous speech, using advanced machine learning. Because differences in language can be so indicative of cognitive differences, we have an opportunity to build unique tools that can effectively monitor speech remotely. I propose to improve and evaluate this technology over the public phone network using a telephony-enabled server accessible by a 1-800 number. This system will automatically receive calls from (and make calls to) individuals, and completely automate speech-based tests, including question-answering, narratives, repeated speech, and random item generation (RIG). There is no telephony research system in the Department (or in the University that we know of) that records speech data over the phone, nor one that engages in human-computer speech interaction. These funds will pay for a Falcon NorthWest computational server. It has 256GB of RAM - no server in our group has this capacity, but it has become increasingly necessary, as our requirements grow with the size of our data, which will need over 2 TB of (backed up) storage. Since machine learning will employ modern ‘deep learning’ methods, it will also be necessary to purchase CUDA-enabled NVIDIA graphics cards, which can accelerate processing. We will use existing phone lines to receive phone calls to our proposed system. However, connecting the phone lines to the computer server requires special telephony hardware. We require F loop-start (FXO) and/or station (FXS) ports on a half length PCI or PCI Express (PCI-e) bus form factor, of the type made available by VoiceTronix. We are budgeting $1376 for this complete solution.