Data Engineer Consultant for ASR (Automatic Speech Recognition)

Terms of Reference for a Data Engineer Consultant for ASR (Automatic Speech Recognition) under the LONGA Project

A. About Farm Radio International

Farm Radio International (FRI) is a Canadian non-profit organization using radio, mobile telephony and other digital tools and services to meet the information and communication rights – at scale – of rural and marginalized communities in sub-Saharan Africa. For rural families in Africa, access to information and communication, e.g. about agriculture, climate change, livelihoods, health and nutrition, are critical needs. FRI is committed to gender equality and inclusion as a cross-cutting priority as well as an end in itself. Our communication solutions are designed to respond to and transform gender relations and the participation of women, youth and marginalized groups at each stage of program design and delivery processes. Obtain more details from the Farm Radio website.

B. Background

With funding from the Wehubit program of ENABEL, Farm Radio International (FRI) is implementing the Longa - Automatic Speech Recognition (ASR) for African Languages (Longa) project. The aim of the project is to transform speech into actionable text with high accuracy and contextual awareness. The project is being developed by FRI with support from CGIAR, leveraging advanced machine learning and natural language processing techniques. For instance, Longa will be used to analyse and respond to the growing number of voice messages that FRI’s radio station partners receive from small-holder farmers through interactive radio programs on our Uliza interactive application. The ‘primary users’ of Longa are the radio station staff - radio presenters or producers, and FRI or other organisations who are gathering voice-based information from farmers/listeners. The ‘end-users’ are the people who listen to the radio programs - the project participants that FRI targets: men and women small-scale farmers. The anticipated outcomes of the project are as follows:

• Improve accuracy for Luganda language and accents recorded via phone call
• Replicate the ASR model for the Bambara language
• Integrate ASR application in FRI’s communication cloud platform for broadcasters

To enhance the performance and reliability of the ASR system, we seek a skilled Data Engineer consultant with expertise in speech recognition technologies. The primary objective of this consultancy is to provide expert guidance and technical support in the development, testing, and optimization of the ASR algorithms. The consultant will work closely with our development team to ensure the highest levels of accuracy, efficiency, and scalability in the ASR system.

C. Scope of Work

The consultant will be responsible for the following tasks:

• Data Management: Collect, clean, and preprocess audio and textual data to train and evaluate ASR models.
• Data Pipeline Development: Design and implement scalable data pipelines to automate the extraction, transformation, and loading (ETL) of speech data.
• Model Training: Work closely with the data science team to prepare data for training, validating, and testing ASR models.
• Performance Monitoring: Monitor and analyze the performance of ASR models, including accuracy, latency, and error rates. Identify and address any data-related issues.
• Integration: Collaborate with software developers to integrate ASR models into web applications, ensuring seamless functionality and optimal user experience.
• Documentation: Create and maintain documentation related to data workflows, model performance, and integration processes.
• Collaboration: Work with cross-functional teams including product manager, UX/UI designers, and data scientists to align data engineering efforts with overall product goals.

D. Deliverables

Data Collection and Management:

• Data collection scripts and tools for gathering speech and text data.
• Clean and labeled datasets for training and evaluation purposes.
• Data storage solutions with appropriate indexing and retrieval mechanisms.

Data Pipeline Documentation:

• Design documents for data pipelines, including flow diagrams and architecture.
• Source code and configurations for ETL processes.

Model Preparation:

• Scripts and processes for preparing and augmenting data for ASR model training.
• Documentation on dataset splits (training, validation, test).

Model Performance Reports:

• Performance metrics reports for ASR models, including accuracy, precision, recall, and F1 scores.
• Analysis and troubleshooting reports on any data-related issues impacting model performance.

Integration Documentation:

• Integration guides and API documentation for incorporating ASR models into web applications.
• Code samples and best practices for interacting with ASR services from web clients.

User Feedback Analysis:

• Reports and insights from user feedback related to ASR accuracy and usability.
• Recommendations for improvements based on user feedback and performance data.

Codebase and Version Control:

• Version-controlled code repository for data processing scripts and pipeline implementations.
• Regular updates and commits to the codebase reflecting ongoing improvements and bug fixes.

Collaboration Reports:

• Documentation of cross-functional team meetings and decisions.
• Progress reports on collaborative projects involving ASR integration.

Testing and Quality Assurance:

• Test cases and results for ensuring data quality and model accuracy.
• Documentation of QA processes and issues encountered during testing.

Training and Support Materials:

• Training materials and documentation for team members on data handling, model integration, and troubleshooting.

E. Duration and Timeline

The consultancy will commence in October 2024 and conclude in August 2025 with a maximum level of effect (LoE) of 80 days. Specific deliverable deadlines will be agreed upon at the start of the consultancy.

F. Reporting

The consultant will report to the Project Manager Nathaniel Ofori and provide regular updates on progress. Weekly meetings will be held to discuss challenges, milestones, and next steps.

G. Elements of application

Interested applicants should submit a technical and financial proposal that includes the following:

• Qualifications and relevant experience of the firm/specific consultant that will be leading the assignment
• Clearly demonstrate experience in Modelling, Fine-tuning, and Deployment of Custom ASR Models
• The general approach that the consultant(s) will take to accomplish the assignment
• Specific tasks/actions that the consultant will undertake to complete the deliverables
• The timeline and work plan
• Proposed budget