DeepMyna Phase 2

Habitat Learn Limited (HLL) provides a personalised learning software, Messenger Pigeon, using its own Automated Speech Recognition("ASR") and generative AI to improve learning outcomes for students.

HLL has 50,000+ UK and North American student-users, and 400 HEI (B2B2C with student as the end user) customers in UK and North America.

DeepMyna builds upon IUK project DeepSpark, an AI toolkit optimising automated speech recognition (ASR) models for higher education, and DeepMyna Phase 1 Feasibility Study.

ASR systems are widely used but can exhibit bias in a multitude of ways - one size does not fit all, leading to inequities, with certain types of speakers being transcribed incorrectly by models that otherwise perform well. Large-Language-Models ("LLM") can help speech recognition better predict words by rescoring the word output from spectrum data analytics, but can introduce domain-specific bias, causing misprediction or ignoring of key information.

DeepMyna builds on a feasibility study completed in May 2023 which completed the following evaluation:

* Bottlenecks identified which limit the adoption of trustworthy AI;
* How DeepMyna's proposed evaluation and customisation model might support acceleration of adoption of trustworthy and responsible Artificial Intelligence (AI) and Machine Learning (ML) solutions;
* SME use cases for accessing new markets and supporting their own product and market development

As part of the feasibility study Habitat Learn assembled a consortium comprising University of Southampton, Avanade Limited, Microlink PC (UK) Limited and Affiniti AI Limited.

DeepMyna is proposed as a world-first bias-driven trustworthy automatic speech understanding (TASU) model evaluation and customisation platform with automated reporting and fine-tuning. Combining ASR and LLM to maximise the understanding of speech content, this regulatory platform will reinforce user-engagement, transparency, and privacy through bias-detection and user-feedback.

DeepMyna will address identified bottlenecks:-

* Lack of transparency in the diversity and representation of training data;
* ASR models are?not explainable;
* Lack of user input for fine-tuning; and
* Lack of privacy protection.

DeepMyna will be based on HLL's datasets compiled from 350,000 hours of lecture recordings with manually curated transcript and summary notes. HLL will extract provenance data for ASR training, such as language, background noise, gender, age, accent, etc, which will be applied to the evaluation matrix, to address the bias issues.