"DAME lays the groundwork for a living interview solution that could support positive investigative outcomes."
Interviews form an important part of any investigative process. They become valuable digital assets that require care and attention in the form of appropriate management, access control, transcription, storage and retrieval.
One government agency was looking to better store and manage interview data securely via a new data platform, with machine learning-based automated transcription.
The challenge
The existing system required administrative personnel to transcribe interviews manually. With a high degree of variation in audio quality and an average interview length of one hour, the process was time consuming and expensive.
An hour interview could take five to 10 hours to transcribe, depending on the quality of the recording. A backlog of transcriptions would quickly build up. Search and retrieval times were also an issue due to legacy storage practices using physical media and warehouses.
The approach
DAME combines organisational-centric design with AWS machine learning capabilities.
The Kablamo-built DAME (Digital Asset Management Experience) was the tool to meet these priorities. The customisation ranged from custom vocabulary to high-level security practices, such as log auditing and encryption.
The initial prototype was delivered in less than two weeks, with user feedback aiding the optimisation of the finalised implementation. AWS Transcribe was tested with real interview data of different qualities and varieties — including accents, slang usage, diction, recording equipment, and file type.
Results
The speech-to-text automation reduced administrative staff's workload by 50% or more for most audio recordings. Machine learning solutions were specially trained to recognise regular phrases and local place names, remove mutters, and apply grammar rules to common phraseology.
Looking forward
This solution goes well beyond transcription efficiencies. It opens up a powerful machine learning-based future that can support investigative outcomes. In the near future, disparate interview data can be securely meta tagged, searched and linked as the platform continues to evolve.
