Actively building|Open for contributors

Tara

Open-source, India-trained voice AI for the 400 million workers who need it most.

Agentic AI for India's Most Common Problems

Building for
01

Wage Math

"আমি ২২ দিন কাজ করেছি ৬০০ টাকা করে। ম্যাডাম দিয়েছেন ১১,০০০। ঠিক আছে?"

"I worked 22 days at ₹600. Madam gave ₹11,000. Is that right?"

Bengali

She's owed ₹2,200 more. Tara will tell her.

02

Time & Day

"आत्ता किती वाजले? आज मंगळवार आहे का?"

"What time is it? Is today Tuesday?"

Marathi

No friend needed. No screen needed. We're building this first.

03

Navigate

"ಮಡಿವಾಳದಿಂದ ಇಂದಿರಾನಗರ 100 ಅಡಿ ರಸ್ತೆಗೆ ಹೇಗೆ ಹೋಗಬೇಕು?"

"How to reach Indiranagar 100 feet road from Madiwala?"

Kannada

Landmarks and bus numbers. How people actually give directions.

+
04

Health Triage

"ল'ৰাৰ তিনিদিন ধৰি জ্বৰ। PHC ১৪ কিলোমিটাৰ। যাম নে?"

"Child has fever three days. PHC is 14 km. Should I go?"

Kamrupi

Which clinic is actually open. Not which one is listed. This agent is in early design.

05

Schemes

"ମୁଁ ଆୟୁଷ୍ମାନ ଭାରତ ପାଇଁ ଯୋଗ୍ୟ କି? ଆଧାର ନାହିଁ।"

"Am I eligible for Ayushman Bharat? No Aadhaar."

Odia

Honest answers. Even when the answer is difficult. We're mapping scheme data now.

06

Rights

"मालकिनले भन्छिन् पुलिस बोलाउँछु। मैले गिलास फुटाएँ।"

"Madam says she'll call police. I broke a glass."

Nepali

The answer is no. Tara will say so. In Nepali. Without hedging.

07

Remittance

"नेपाल पइसा पठाबय के सबसँ सस्ता तरीका कोन?"

"Cheapest way to send money home to Nepal?"

Maithili

Real comparison. Informal channels included. Corridor data collection underway.

How it works

01

Missed call

She dials a local number. Tara calls back. Zero cost to her.

02

She speaks

Her language, her dialect. No menus, no commands, no prompts.

03

Agent routes

Intent is detected. The right domain agent activates automatically.

04

Voice reply

Spoken answer in under 8 seconds. No screen. No text.

Tech Roadmap
USSD gatewaySIM toolkitWhatsApp voice notesKaiOS (Jio Phone)
400M+
Informal workers in India

No AI product is being built for them. We are.

0
Literacy required

Voice in. Voice out. No screen.

7
Agents in development

Wages. Time. Health. Rights. Navigation. Schemes. Remittance.

Why We're Building This

She owns a smartphone. She calls someone to ask what time it is.

A domestic worker from Nepal, living in Bengaluru. Illiterate. Undocumented. She cannot read the numbers on her own screen.

So she built her own system — she dials a friend, asks the time, acts on the answer. A perfectly functional information retrieval system, engineered around her reality.

But at 4 AM, nobody picks up. And the question isn't always the time. Sometimes it's "has madam paid me right this month?" Sometimes it's "my child has been sick three days — should I spend \u20B9300 on an auto to the hospital?"

We are building Tara so she gets an answer at 4 AM too.

You know someone like her. She works in your building. She watches your gate. She raised your children while raising her own.

Help us build the thing that finally reaches her.

Under the Hood

Indian voices. Indian data. Indian hardware.

Tara is not a wrapper on a Western LLM. We are training a ground-up Indic language model on spoken dialect data being collected across India. The model is learning the language as it is spoken, not as it is written.

Architecture
Decoder-only transformer, being distilled for spoken Indic languages
Parameters
Target: 1.5B (edge) / 7B (server) — architecture experiments in progress
Attention
Grouped-query attention (GQA) with RoPE positional embeddings for long-context voice turns
Quantization
INT4/INT8 mixed-precision via GPTQ + ONNX Runtime for Jetson deployment
Vocab
Custom 48K BPE tokenizer trained on spoken Indic corpora — phoneme-aware, code-switch friendly
Context
4096 tokens (voice turn context) — sufficient for multi-turn spoken dialogue
Training Data
Spoken dialect corpora being collected across India — bus stands, union offices, helplines
Data Pipeline
Whisper-large-v3 transcription → human dialect correction → alignment → training pairs
Languages
Bangla, Nepali, Assamese, Odia, Bhojpuri, Maithili, Sylheti, Kamrupi
Register
Colloquial, code-switched, street — not literary
ASR / TTS
Bhashini-native speech pipeline + custom dialect fine-tuning in progress
Latency Target
End-to-end voice response under 8 seconds (ASR + inference + TTS)
Edge Target
NVIDIA Jetson Orin NX (8GB) — on-device inference, no cloud dependency
Fallback
Server-side 7B model via IVR for complex queries or low-end devices
Deployment
Missed-call IVR, USSD, SIM toolkit, WhatsApp voice, Android overlay, KaiOS
License
MIT — fully open source
Evaluation
Zero-literacy task completion rate, not MMLU — measured via field callbacks
IN

Being trained in India. On Indian data. For Indian realities.

We are collecting the training corpus directly from the communities Tara will serve — recording with consent through domestic worker unions, ASHA worker networks, and migrant labour organisations. This data does not exist on the internet. We are creating it. And we need researchers, linguists, and field partners to help.

Tara system architecture

ASRIntent RouterDomain AgentsTTSIVR / Android Overlay
NVIDIA Jetson Orin NXBhashini ASR/TTSPyTorchHuggingFaceONNX RuntimeUSSD/IVRAndroid OverlayMIT License

We're looking for

Research Output

Publications

Indic ASR

Low-Resource Automatic Speech Recognition for Eastern Indo-Aryan Dialects: A Contrastive Pre-Training Approach

SARAL Working Paper, 2026

Voice LLM

Dialect-Aware Spoken Language Modelling for Non-Literate Populations: Architecture and Evaluation on Bangla-Sylheti Code-Switching

Submitted to ACL 2027

Edge Inference

Sub-8-Second Voice Agent Inference on ARM Cortex-A53: Quantization and Pipeline Scheduling for Feature Phone Hardware

SARAL Working Paper, 2026

Agentic Systems

Intent Routing Under Ambiguity in Voice-Only Multi-Agent Systems: A Decision-Theoretic Framework for Low-Bandwidth Channels

SARAL Working Paper, 2026

Data Collection

Participatory Corpus Construction with Non-Literate Communities: Consent Protocols, Compensation, and Annotation Quality

SARAL Working Paper, 2026

TTS

Prosodically Faithful Text-to-Speech for Indian Domestic Worker Registers: Evaluation Beyond MOS Scores

Under Review, 2026

Wage Estimation

Structured Information Extraction from Conversational Wage Queries: A Slot-Filling Approach for Informal Labour Arithmetic

SARAL Working Paper, 2026

Evaluation

Beyond BLEU: Task Completion Metrics for Voice AI Serving Illiterate Users in High-Stakes Information Retrieval

SARAL Working Paper, 2026

Section 8 Non-ProfitNVIDIA InceptionOpen Source (MIT)Published: Wiley, HuggingFaceMade in Calcutta

This is being built in the open. Join us.

Whether you're a developer, a linguist, a funder, or someone who simply believes this should exist — we want to hear from you. Leave your name. Tell us how you'd like to help.

No commitment. Just a conversation to start.

Looking for a specific kind of partnership?

Get involved