Highlight
The Digital India initiative by the Government of India, combined with smartphone penetration in all sections of society, has led to the need for AI in building inclusive digital systems for critical public services like finance, healthcare and agriculture. The R&D team in Hitachi India is actively engaged in researching speech (the most natural medium of communication) as a means to democratize access to those essential public services, particularly in banking.
This article discusses three critical aspects of speech-based financial inclusion - voice-based authentication, vernacular speech recognition, and embedding of automated speaker verification in edge devices. This latter is helpful for spilt-second inferencing in high-latency and low-bandwidth situations where a miniaturized speech comprehension engine is available locally in a smartphone app, which may even be offline.
We also discuss the development of our AI model architecture from scratch with focus on understanding of Indian vernacular languages; appropriate quantization of the neural network model for miniaturized footprint; and the challenges related to accuracy. Finally, it touches upon the company’s approach of building a relevant dataset, focusing on connected numbers, a key need for “financial amount” articulation in speech-based transactions.
