Documente Academic
Documente Profesional
Documente Cultură
In 2017, the UN Department of Economic and Social Affairs reported a world population of 7.6
billions. By the year 2050, this figure is expected to leap to 9.8 billions.
Africa will represent more than 50% of that growth with a 2017 population of 1.3 billion projected
to reach 2.5 billions. Africa’s population explosion is concentrated in areas in which corruption
belies the very potential for scalable and sustainable development. The Sela Protocol efficiently
and effectively connects pools of capital with entrepreneurs that need them most and create
feedback loops connecting beneficiaries with funders to ensure transparency, accountability.
The Sela Portal enables a open communication across stakeholders throughout a project’s life
cycle powering the development of a trust economy.
A) Designing for the next billion internet users: the need for voice-first interfaces
In pilot tests conducted in Spring 2018, it became apparent that some of our evaluation agents
encountered difficulties using our text-based interface to upload project claims, answer
questions and spell passwords. Further research proved that this is not an isolated issue as
close to 40% of the adult population in Nigeria is illiterate.
Google is leading the way in India to design interfaces for such population with their “next billion
users” initiative. In the blog post, they note that “many of India’s new internet users favor
listening and speaking over reading text.” Our team made similar observations in Nigeria as
illiteracy prevents users from entering password information, answer questions and manage
their funds. New approaches are needed to design for such users.
We will now describe four sample projects inspired by our learnings and needs.
4) Projects
A) Project 1: Combined speaker & speech recognition for data submission (Speaker &
Speech Recognition)
For data submission, the typical workflow for an evaluation agent is receiving a notification or
trigger prompting them to go on a given site which they will be familiar with and answer a series
of questions.
For example, a woman living in K-Dere in the Niger Delta can be prompted to go inspect a
school or road construction, take pictures, indicate which task the media corresponds to and
answer a series of follow up questions which at the beginning will include:
1) Yes/no questions
2) Numerical answers
3) Narrative answers
In that process, instructions can either be given in English or local language and answers can
also be given in English and tribal languages.
For English questions and answers, field experiments with readily available speech recognition
engine can be made by the Sela team to test their error rate.
The scope of this project includes creating combined speaker & speech recognition engine to be
deployed on cheap Android smartphones for specific numeric and binary keywords spoken with
non english accents as well as local languages such as :
● Pidgin
● Yoruba
● Igbo
● Hausa
● Ogoni
The scope of the project will include:
● Exploration of existing data sets for our needs
● Approaches needed to collect additional data
● Quantifying added value of directly learning from the data captured as well as
approaches such as transfer learning to combine big and small data from other
languages
● Field testing of the speech recognition engines in collaboration with the Sela team
● Quantifying robustness and ease of spoofing of speaker recognition models