Scaling Voice AI for Agriculture
100+ Hours
Ground-truth voice data transcribed
15 Languages
Supported across India & Africa
800K Downloads
Farmer.Chat adoption growth
High-quality multilingual transcription enables domain-specific speech models for smallholder farmers across low-resource languages..
Challenge
Digital Green is building Farmer.Chat, an AI-powered assistant designed to support smallholder farmers with real-time agricultural advice. The system enables users to submit queries via voice, text, or images in native languages, addressing diverse crops and regional contexts.
However, scaling this capability exposed a fundamental limitation: existing speech-to-text models perform poorly on low-resource languages and domain-specific agricultural vocabulary. Over 70% of queries are in native languages, with ~46% coming through voice and image inputs, making accurate speech recognition mission-critical.
Generic ASR systems struggled with:
- Local dialects and pronunciation variability
- Agriculture-specific terminology (crop names, diseases, inputs)
- Real-world audio conditions (noise, pauses, low-cost devices)
Even minor transcription errors could lead to incorrect or irrelevant recommendations, undermining trust and usability.
“Even the best speech models fail for agriculture-specific queries in low-resource languages.”
Solution
- Built a scalable human-in-the-loop transcription pipeline using iMerit’s ANGO platform
- Delivered high-quality native-language voice transcription and translation
- Enabled domain-specific ground truth dataset creation
- Supported model benchmarking and error analysis workflows
- Established continuous data delivery and QA feedback loops
The workflow was deployed on iMerit’s ANGO platform, enabling:
- Task distribution to trained annotators
- Multi-layer quality assurance workflows
- Continuous batch processing (weekly/monthly ingestion cycles)
- Structured output delivery for downstream ML pipelines
iMerit partnered with Digital Green to design and operationalize a high-quality data pipeline for speech AI development in low-resource agricultural contexts.
At the core of the solution was human-annotated transcription, where native-language experts transcribed farmer voice queries into precise ground truth data.
This process focused on capturing linguistic nuance, regional dialects, and agriculture-specific vocabulary—elements often missed by automated systems. In addition to transcription, iMerit supported select translation tasks, bridging gaps where automated translation pipelines underperformed.
This curated dataset became foundational for Digital Green’s internal ML workflows, including:
- Benchmarking multiple speech-to-text models against real-world data
- Identifying systematic error patterns (e.g., misrecognized crop names)
- Enabling targeted fine-tuning strategies such as:
- Contrastive learning for similar-sounding terms
- Curriculum learning based on utterance complexity
- Weighted loss functions for high-impact vocabulary
Unlike synthetic datasets, this data reflected actual user behavior, including noise, pauses, and unstructured speech—making it significantly more valuable for production model improvement.
Result
The collaboration enabled Digital Green to build one of the most robust domain-specific speech datasets for agriculture in low-resource languages, unlocking measurable improvements in model evaluation and optimization.
Key outcomes include:
- Creation of 100+ hours of high-quality Hindi agricultural voice data, with expansion into additional languages underway
- Development of a reliable benchmarking framework to evaluate ASR models in real-world conditions
- Identification of critical failure modes in existing speech models, particularly in domain-specific terminology
- Acceleration of fine-tuning strategies targeting high-impact transcription errors
These improvements directly support Farmer.Chat’s ability to deliver accurate, context-aware recommendations to farmers at scale. The platform has already processed 6 million+ farmer queries, with growing adoption of voice interfaces.
Looking ahead, Digital Green plans to:
- Deploy domain-adapted speech models in multiple languages (Hindi, Amharic, Hausa)
- Open-source datasets and benchmarking tools to advance the broader AI ecosystem
- Expand into new geographies across Africa, Asia, and Latin America
“The ground truth data enables us to benchmark models, understand errors, and design targeted strategies to improve performance.”
