SamarthVaani SamarthVaani (meaning "empowered voice" in Hindi) is a voice-enabled, multilingual AI assistant built for persons with disabilities in India — specifically designed to help them understand and access their legal rights and government benefits under the Rights of Persons with Disabilities (RPWD) Act, 2016. The system is built for a population that is often underserved by digital tools: people with physical, intellectual, sensory, or mental disabilities who may speak regional Indian languages, have low literacy in English, or struggle to navigate complex bureaucratic information on their own. At its core, SamarthVaani combines a Retrieval-Augmented Generation (RAG) pipeline — built on FAISS vector search and sentence-transformer embeddings — with Sarvam AI's Indian-language speech APIs to let users ask questions by voice or text in any of 11 Indian languages (Hindi, Tamil, Telugu, Marathi, Bengali, Kannada, Malayalam, Gujarati, Punjabi, Odia, and English), and receive answers spoken back to them in their own language. The knowledge base is sourced from curated government documents including the full RPWD Act 2016 legal text, a plain-English disability rights manual, the Ministry of Social Justice's Compendium of Schemes, structured data on 8+ government schemes (ADIP, DDRS, SIPDA, NHFDC), 7 national disability institutes with contact details, and 6 ready-to-use application letter templates — all indexed together so that a villager or caregiver can simply ask "मुझे व्हीलचेयर कैसे मिलेगी?" ("How do I get a wheelchair?") and receive a direct, spoken answer with eligibility criteria, documents needed, and the office to approach.
Who We're Building This For SamarthVaani is built for persons with disabilities in India and their caregivers — a population of over 26.8 million people (Census 2011, likely undercounted) who are legally entitled to rights, schemes, and benefits under the RPWD Act 2016, but in practice cannot access them because the information is:
Written in legal English they cannot understand Scattered across dozens of government portals Inaccessible to those who cannot read, type, or see More specifically, the primary users are:
User Type Problem They Face Rural PwD with low literacy Cannot read government scheme PDFs; needs spoken answers in Hindi/regional language Deaf/hard-of-hearing person Needs text-based access to rights information without phone calls Parent/caregiver of a child with intellectual disability Doesn't know what schemes exist or how to apply Visually impaired person Needs voice-first interaction; screen readers don't help with bureaucratic navigation NGO field worker Needs quick factual lookups when counselling beneficiaries in villages What Could Go Wrong
- Hallucinated or outdated answers (highest risk) The RAG system retrieves real chunks but the extractive synthesizer stitches sentences from different contexts. Without an LLM to reason over them, the answer can be grammatically correct but factually misleading — e.g., mixing eligibility from one scheme with benefits from another. If a person acts on wrong information about disability certificate thresholds or income limits, they could be wrongly denied or wrongly encouraged to apply.
Mitigation: Always show source citations; add a disclaimer that users should verify with the District Disability Officer.
- Language quality degradation through the translation chain Every non-English query goes through: User language → English (Sarvam) → RAG retrieval → English answer → User language (Sarvam) → TTS. Each step introduces errors. A question asked in Marathi may lose domain-specific terms like "बेंचमार्क दिव्यांगता" (benchmark disability) after round-trip translation, producing a subtly wrong answer.
Mitigation: Test with native speakers of each supported language; use mode: "formal" in the Sarvam translate API (already done).
- Voice STT failure for disability-affected speech The very users this app serves — people with speech disabilities, cleft palates, cerebral palsy affecting speech, or heavy regional accents — are the most likely to have their voice input mis-transcribed by the STT model. A person with spastic dysarthria asking about cerebral palsy benefits may get a blank transcript.
Mitigation: Always show the transcribed text so users can correct it; keep the text input tab as a full fallback.
- API dependency and cost at scale Sarvam AI's TTS/STT/Translate APIs are rate-limited and paid. If this gets deployed publicly and usage spikes, API costs could become prohibitive, or rate limits could silently break voice features — with no clear error shown to users.
Mitigation: Add per-session caching of TTS audio; show clear error messages when API limits are hit; budget API costs before public deployment.
- Data staleness Government schemes change: income thresholds increase, new schemes are added (e.g., PM-DAKSH, Unique Disability ID expansion), old ones are merged or renamed. The CSV/PDF knowledge base was last updated in 2019–2023. A user asking about current UDID application procedure may get a deprecated 5-step process.
Mitigation: Tag every chunk with a last_updated field; build an update pipeline that re-ingests when new ministry circulars are released.
- Trust and misuse The app could be mistaken for an official government service. A user might believe a generated application letter is legally sufficient without cross-checking. The application templates use placeholders like [UDID_NUMBER] — if a user submits an unfilled template, their application will be rejected.
Mitigation: Add a prominent disclaimer — "SamarthVaani is an information tool, not an official government service" — and validate template fields before generating letters.
Track - 4 Governance
Built With
- ai
- llm
- rag
- streamlit
- stt
- tts
Log in or sign up for Devpost to join the conversation.