Sabaq: Voice Tutor for Urdu-English Classrooms

Sabaq is a voice-first learning assistant for Pakistani classrooms where teachers speak Urdu but keep technical terms in English. In ordinary speech-to-text, terms such as “overfitting” or “machine learning” may appear in Urdu or Devanagari script, making lecture transcripts difficult to search and reuse. Sabaq transcribes uploaded or live lectures with AssemblyAI, restores observed technical-term variants to consistent English spellings, and turns the corrected transcript into summaries, key points, and practice questions. Students can also ask the lecture questions by voice; answers are constrained to the transcript and the app flags topics the teacher did not cover. Gemini generates study material, with Groq as fallback. We evaluated Sabaq on 7 scripted lectures, 4 speakers, and 49 scored transcripts. In our tests, setting batch transcription to Urdu increased searchable technical-term recall on new speakers from 16% to 79%. The evaluated live path reached 64%, while the AI/ML glossary improved one harvested lecture from 23% to 100%. Median live lag was 1.4 seconds. The README documents the limits, failures, and reproducible evaluation.