REQUEST FOR PROPOSAL (RFP)
Speaker Attribution Data Annotation of ORF Voice Recordings
Closing Date for Submission
15th September, 2026, at 23:59 PM
LEHS is a charitable organization in India whose purpose is to offer basic health and education for the poor. LEHS in furtherance of charitable objectives through its flagship programs, Wadhwani AI, which aims to build equitable and sustainable systems by making quality primary healthcare available and accessible to the underserved population, and to bring the benefits of modern AI technology to underserved populations by building and deploying AI solutions for social impact across domains such as healthcare, agriculture, governance, and education in India. LEHS aims to promote the integration of technologies, particularly in emerging domains like artificial intelligence and innovations into the Indian mainstream primary healthcare, education, and agriculture systems through a partnership with the State and National Government, apex institutions, international agencies, and private sector partners e.g. innovators, social enterprises and other ecosystem contributors in line with its stated objectives for the betterment of society particularly focusing on projects of national and social significance. LEHS also undertakes rigorous monitoring, evaluation, and learning (MEL) activities to assess the impact, usability, and scalability of different programmatic interventions.
Wadhwani AI, a unit of LEHS, focuses on developing, deploying, and evaluating artificial intelligence solutions to address critical social challenges in India, particularly in domains such as healthcare, agriculture, and education.
A significant number of students in Indian public schools (Grades 1–8) lack foundational reading skills, impacting their academic progression across all subjects. This challenge is compounded by the absence of scalable, data-driven tools for assessing reading in diverse languages and for identifying specific proficiency levels to enable targeted interventions.
In response, LEHS (Wadhwani AI) has developed an AI-powered tool that automates the assessment of foundational reading skills. The tool leverages highly contextualized and fine-tuned Automated Speech Recognition (ASR) models to analyze word and paragraph reading, generating key performance metrics and identifying specific reading miscues. Our solution also applies an intelligent layer to cohort students into reading levels, providing teachers with actionable insights for planning targeted classroom interventions.
To date, LEHS (Wadhwani AI) has assessed over 8 million students in Gujarat and Rajasthan for the Gujarati and Hindi languages. When the assessments are done in a classroom environment, there are many issues with the quality of the audios that teachers use for child’s assessment as ASR catches multiple speakers including adult reading in place of a child. We want to improve our mechanism to detect such speakers and train models that can perform well when people are speaking over each other, especially adults.
This RFP is issued to solicit bids from experienced data annotation agencies to generate highreliability ground truth data for speaker attribution within Oral Reading Fluency (ORF) assessment audios. This dataset will support the development of models capable of detecting multiple speakers, specifically substitution (an adult reading in place of the child), prompting (an adult providing assistance to the child) and overlap (an adult speaking over child reading).
PROPOSAL SUBMISSION GUIDELINES
Your proposal must include:
1. Technical Proposal
- Understanding of Scope, Deliverables, and Problem Statement: Demonstrate understanding of the problem we are solving with these annotations, the full scope of work, and the expected deliverables.
- Process and Tools
- Annotator selection criteria
- Equipment and listening environment standards
- Annotator training process: how annotators are trained to understand each type of annotation and the differences between them
- Annotation tooling used
- How annotator fatigue is managed
- How annotation quality is defined and what mechanisms are used to monitor annotation quality continuously during the project.
- How underperforming annotators are identified and addressed
- Team Composition and Qualifications: Details of the team assigned to this project, their relevant experience, and role allocation.
- Quality Assurance Mechanisms:
- Remediation process: what happens when annotation accuracy falls below acceptable levels, including the process for how affected data is re-annotated.
- Reporting cadence: how frequently the partner will communicate with us during the engagement (daily, weekly, or milestone-based).
- Reporting content: what specific metrics or summaries will be shared with us on a regular basis (e.g., annotation throughput, flagged files, annotator performance summaries) and in the pilot and final reports.
- Data Policy: partners must describe their data handling policy with respect to our data, including storage location, access controls, retention period, deletion process upon project completion, and any subcontracting or third-party access to the data.
2. Financial Proposal
- Detailed cost breakdown
- Schedule of Payments Linked to Milestones Achieved Applicable taxes and contingencies
3. Organizational Profile o Relevant project experience o Sample of similar work
4. Signed NDA: Partners are expected to download and update the Organisation Name, registration no., Address in the NDA here, provide Authorised Signatory and attach signed NDA with your proposal. Without Signed NDA, Proposal will be invalidated.
5. Annotation of Sample Audios: Partners are provided with a sample set of audio files via Google Drive. Along with their proposals, they must return completed annotations for this sample set as part of their proposal, following the definitions provided above. The format for the annotations is provided in the Annotation Format section. Without Sample annotations in required format, Proposal will be invalidated.
Sample Evaluation Criteria
Partners are required to submit completed annotations for the sample audio set provided. These annotations will be evaluated on annotation accuracy annotations calculated against golden ground truth prepared internally. This measures how closely the partner's annotations align with the expected annotators, and is an indicator of how well annotators have understood and applied the metric definitions and score descriptors.
Criteria for Application
Applicants must meet the following criteria.
Mandatory:
- Demonstrated team capacity to annotate audio clips based on metrics
- Clear and demonstrable quality check process
- Submission of evaluation samples (Proposal without sample annotations will be rejected)
Preferred: Prior experience in annotating multilingual audio data, including languages such as Gujarati, Hindi, English, Odia, Telugu, etc.
Submission Details
- All proposals to this RFP must be received no later than <Sept 15th>. The proposal should be submitted only through e-mail in PDF format addressed to The Procurement Team at the below-given e-mail id: rfp.lehs@wadhwaniai.org. The email's subject line must contain the reference number and title of the RFP: Speaker Attribution Data Annotation of ORF Voice Recordings
- Any proposals received by Wadhwani-AI after the deadline for submission of proposals prescribed in the timeline of this document are liable to be rejected.
For detailed information, please check the complete version of the RFP attached below