VoiceCore
"Build A Voice Agent That Sounds Indian And Thinks Clearly."
Build a Hindi/Hinglish voice agent that can understand Indian speech, answer questions from a knowledge base, respond in a natural Indian voice, maintain conversation context, and gracefully handle questions it cannot answer.
The Problem Statement
Most voice agents are built for clean English, Western accents, and controlled environments. Indian users need AI that understands Hindi, Hinglish, regional accents, noisy speech, and real conversations. The goal is to build a practical Indian voice AI agent that is useful, fast, natural, and grounded in knowledge.
What You Need To Build
Start with the core requirements, then layer advanced features and bonus innovations to stand out.
Must-Have
Minimum features required for a valid submission.
- Browser microphone input
- Push-to-talk or VAD-based speech detection
- Hindi/Hinglish/Indian English ASR transcription
- LLM-based response generation
- Knowledge base Q&A from PDF/text/URL
- Text transcript display
- Indian voice TTS output
- 3-turn conversation memory
- Fallback response when answer is not known
Should-Have
Features that make the prototype stronger.
- Same-language response
- Source citation from knowledge base
- Voice persona selection
- Latency measurement
- 5-turn memory
- Session transcript
- Better handling of Hinglish
Bonus
High-impact additions for standout teams.
- Real streaming pipeline
- Low latency response
- Regional Indian language support
- Offline/edge mode
- Voice speed/tone adjustment
- Cost per 1000 interactions estimate
Suggested User Flow
Initialization
User uploads a PDF menu or company policy to create the knowledge base.
Interaction
User clicks 'Talk' and asks a question in Hinglish.
Processing
Agent transcribes audio, queries the RAG vector store, and generates a response.
Synthesis
Agent reads out the response in a natural Indian accent.
Expected Deliverables
- Working prototype
- GitHub repository
- README with setup instructions
- Demo video (max 3 minutes)
- Architecture diagram
- Dataset/API explanation
- Known limitations
Judging Criteria
Suggested Tech Stack
* These are suggestions only. You are free to use any modern stack you prefer.
Rules & Compliance
The agent must not hallucinate facts. If it does not know an answer or cannot find support in the knowledge base, it must clearly say so and suggest a useful next step.
- Use original work
- Do not copy copyrighted assets
- Do not use unauthorized scraping
- Do not collect private data
- Mention all APIs/data sources
- Open-source allowed with credit
- AI tools allowed, but explain usage
VoiceCore track FAQ
Direct answers about the challenge scope and expected prototype.
What is VoiceCore?
VoiceCore is an XE Labs Build Sprint 2026 challenge track for a practical Indian voice AI agent. The prototype should understand Hindi, Hinglish, or Indian English speech, answer from an uploaded knowledge base, speak through a natural voice, remember recent turns, and handle unsupported questions without inventing facts.
What should teams build in VoiceCore?
Teams should build a browser-based voice agent with microphone input, speech transcription, grounded knowledge-base retrieval, LLM response generation, transcript display, text-to-speech output, and at least three turns of conversation memory. The agent must provide a clear fallback when its knowledge base does not support an answer.
Which languages should VoiceCore support?
The required scope is Hindi, Hinglish, or Indian English, with a strong emphasis on realistic Indian speech. Teams may support additional regional languages as a bonus. The experience should make its actual language coverage clear and should not claim reliable support for accents or languages that were not tested.
Ready to build this track?
Strong submissions and finalists may be invited for paid internship interviews to continue production development under XE Labs.