LocaleNLP

LocaleVoice is a voice-first agricultural assistant built by LocaleNLP to make agricultural information more accessible to Moroccan farmers who communicate in Moroccan Darija. Many digital agricultural tools assume users can comfortably read and interact in dominant written languages. This creates a language and accessibility barrier for farmers who would benefit from asking practical questions in their own everyday speech. LocaleVoice lets users record or upload a voice question, review and correct the transcript, provide additional context when needed, and receive an agricultural response grounded in a curated knowledge base. When the available sources do not support an answer, the system asks for clarification or abstains rather than inventing agricultural advice. The prototype uses Meta's Muse Voice Transcribe model for speech transcription and Muse Spark for evidence-grounded response generation. Its current knowledge base contains selected Feedipedia reference material, and its retrieval and conversation pipeline includes source attribution, insufficient-evidence handling, and multi-turn clarification. The application is built with Python and Streamlit. The current engineering baseline includes 139 passing automated tests and a 6/6 offline evaluation of retrieval and evidence-gating scenarios. LocaleVoice is an early prototype, not a production agricultural advisory service. Moroccan Darija speech accuracy, browser-based end-to-end behavior, human language review, and farmer usefulness still require validation. Our next step is to test with Moroccan Darija speakers and farmers, collect consented speech samples, and benchmark recognition quality. Our goal is to make useful AI-powered agricultural information accessible through the language farmers actually use.