Large language models such as ChatGPT have recently undergone a step-change in their user-perceived efficacy, and as such they have caught the attention of journalists and then the imagination of the general public. These A.I. systems appear to deliver responses to user-posed questions that are both informative and delivered with substantial expertise. The application of this nascent, yet extraordinarily useful, technical breakthrough across whole ranges of industries is only starting to become apparent.
Large language models are effective since they capture and codify, in their models, significant portions of all human knowledge, through the process of ingesting the entirety of the World-Wide-Web. The resultant models are staggeringly large (for example, ChatGPT3 contains some 175 billion parameters) and thus, these models take months to train and use a great deal of power to run.
It is our intention to build a Chat-GPT like large language model A.I. system that utilises a quantum calculation engine. The system shall employ a fundamentally different processing, model and systems architecture to current systems, leading to a step change in the efficiency and power of the resultant system. We shall build software to run such a system and demonstrate its increased efficiency in both training models, which shall be much more energy efficient during the training process, and in speed of execution and energy use when in operation. These benchmarks shall be performed against the requirements of the NLP and voice processing industry as understood by the consortium.
"Insurance fraud costed the UK £3B in 2017, increasing 8% since 2015, equating to £10,400 per fraudulent claim. This causes an increase of £50 per policy. Current manual processes to identify illegitimate claims are repetitive, time-consuming, inefficient, non-user friendly with no consideration for vulnerable customers. This presents fraud identification as a primary target for developed AI capabilities.
The public are becoming increasingly aware of voice activated AI following the release of Amazon's Echo: a market leader in the 'voice-controlled home'. Resultant scientific research surrounding this is focusing on adding intelligence to spoken commands for real-time information.
Current advanced emotional AI technologies capture (non-verbal) human expressions via computer vision, voice analysis and/or biometric sensors. These lack processing speed and are unable to understand both vocalics and linguistics. No solution currently combines artificial intelligence and voice technology for a true conversation with full explainability of its decision.
Addressing this gap, Intelligent Voice aim to develop an AI software that detects and interprets emotion and linguistics from voice. In collaboration with behavioural analysis experts at Strenuus and deep neural networks academics at University of East London, this project aims to develop a vocal AI technology for credibility/vulnerability assessment, key word spotting, in-call behavioural guidance and transparency of the decision-making process, trialled by an insurance contact centre during live claims handling. This will offer a breakthrough technology for the anti-fraud sector, simultaneously providing unique expertise to the UK in deep neural networks and AI with cross-sector potential.
All partners are well placed to exploit this opportunity: Intelligent Voice have an existing global client-base for their voice recognition/transcription software in the financial/insurance services market, Strenuus completed Proof-of-Concept studies with their linguistic algorithms to assess credibility, successfully identifying deception, UEL has demonstrable expertise in data analytics, machine learning architecture and AI explainability. Insurance software providers are in place as in-kind contributors for exploitation.
With InnovateUK support, a 30-month programme of Industrial Research and Experimental Development is required to develop behavioural analysis algorithms, explain the neural networks of the system, integrate the solution to ICE's existing platform and trial the technology in a controlled environment. Project success will support commercialisation by 2021, to establish:
- Intelligent Voice at the forefront of the speech recognition software market, poised for significant growth;
- Strenuus as a leader in behavioural analysis with novel audio processing algorithms;
- UEL as a UK leading university with state-of-the-art explaining of DNNs."
Knowledge Transfer Partnership
To develop and embed expertise in emotion recognition from speech and other modalities using state of the art machine learning techniques including deep learning and kernel machines.
Knowledge Transfer Partnership
To enhance video conferencing by combining Augmented Reality, Automatic Speech Recognition, Natural Language Processing and GPU-accelerated cloud processing to provide a more natural communication with each meeting participant’s avatar superimposed into the user’s real-world view.
The project will enable the processing of fully encrypted telephone data to facilitate secure voice conversations
in the cloud, or locally, such that only the end-user has access to the results. A user will be able to store, search
and convert to text any audio of telephone file in a way that cannot be intercepted or interfered with by a
cloud provider, or any third party in the transmission chain. Currently, a user has the choice of fully encrypting
their voice data for transmission and storage, or transmitting it "in clear". However, each choice involves a
compromise: The secure communication cannot be searched or processed using the processing power
available in the cloud whereas the unencrypted communication can be processed into text, or made searchable
for mass off-site storage but this compromises the privacy of the communication.
This project will adapt techniques originally applied to genomics, and re-interpret them to solve one of the
greatest data privacy challenges to-date: How can you keep control of privacy, while leveraging the power of
the cloud for search and storage?
GRD Development of Prototype
Fraud and market abuse (insider dealing and market manipulation), have a negative effect on
all parts of the economy and are a significant problem in the financial services sector. In order
to address market abuse, the Financial Services Authority (the financial regulator in the UK)
has introduced regulations requiring financial firms to store all data from communications
such as emails, Instant Messaging, texts and recordings of phone calls, for a minimum of 6
months. In order to process and analyse this data to identify fraud and market abuse,
compliance & monitoring platform tools are used; however, existing tools are limited in their
ability to handle and analyse voice data from phone calls effectively.
The main methodologies currently used to handle and analyse voice data (Speech recognition
and Voice Data Mining software) are not capable of identifying separate speakers or grouping
voice data to similar unknown speakers; the sentiment/emotional content of the voice remains
undetected or lost; and responses to regulatory requests can take up to 3 days, delaying
investigations into financial fraud and market abuse.
To address these limitations and to take advantage of a market estimated at $2.8bn, CHASE
ITS Limited, seek to develop a pre-production prototype of an advanced platform
“Accelerated Voice” (AV), a next generation compliance tool for the financial market to
efficiently and accurately process voice data associated fraud and market abuse. AV
encompasses a number of challenging technical innovations involving the development of
multiple advanced functionalities which aim to detect emotion, separate speakers on phone
calls and collate unknown callers, and increase processing capabilities. This project is both
market and customer driven and if successful will offer a unique solution to the market place,
benefitting the financial service sector, through more efficient and transparent compliance
monitoring and an overall reduction in fraud and market abuse