What if a farmer could simply pick up a phone, speak naturally in Malayalam or English, and ask:
“What is today's market price?”
or
“It's going to rain tomorrow. Should I spray my crops today?”
or even:
“My crop is getting damaged. Can someone help me?”
No typing. No complicated menus. Just a conversation.
That was the idea behind Farm & Field, the voice agent I built during the 10 Days of Voice Agents — VoiceForBharat Edition challenge by Murf AI.
Over these 10 days, I went from building a basic voice conversation to creating an agricultural voice assistant with memory, real-time tools, outbound calling, human escalation, call analytics, multilingual conversations, and a specialist agent.
I built Farm & Field, a voice assistant designed to help farmers with everyday agricultural questions.
Farm & Field can:
The AI should not try to do everything itself.
If it needs current information, it uses a tool.
If the problem needs specialized knowledge, it hands the conversation to a specialist.
If the situation is beyond what the AI can safely handle, it asks a human for help.
A voice agent is basically four main components connected together:
Speech-to-Text (STT) — converts the farmer's speech into text. I used Deepgram.
LLM — understands the conversation and decides what to do. I used Google Gemini.
Text-to-Speech (TTS) — converts the response back into voice. I used Murf Falcon.
Real-time transport — carries the audio between the user and agent. I used LiveKit.
On top of this, Farm & Field has:
🇮🇳 Malayalam, English and Manglish support
Farmers don't always speak in one language.
For example: "Ente rubber crop-il leaves yellow aakunnundu."
The agent can recognize Malayalam/Manglish conversations and adapt the TTS voice accordingly.
This made the conversation feel more natural for Indian users.
Murf Falcon for fast voice responses
I used Murf Falcon for TTS because latency is very important in voice conversations.
A slow response makes a voice agent feel like a chatbot that is reading messages aloud.
With a faster TTS response, the conversation feels much more natural.
Memory with consent
Farm & Field can remember useful information about returning farmers, such as:
But the agent asks for permission before saving information.
This was important because memory should be useful without automatically storing everything a user says.
Tools instead of guessing
The agent doesn't rely on the LLM to guess live information.
For example, when a farmer asks: "What is today's rubber price?"
the agent uses the market-price tool.
The same approach is used for weather information.
This makes the system more reliable.
Human escalation
The agent can recognize situations where AI should not be the final answer.
I implemented escalation for cases such as:
The farmer gives consent before a human-help request is created.
Specialist agent
I created a separate Crop Problem Specialist for crop disease and pest-related conversations.
The main agent can hand the conversation to the specialist without making the farmer repeat everything.
The specialist also has a different voice, making the handoff clear to the user.
Outbound calls
Farm & Field can initiate calls for specific alerts, such as:
I used LiveKit SIP for the outbound calling workflow.
Call analytics
The system also tracks call information such as:
This helped me understand what happened during conversations instead of treating every call as an isolated interaction.
Repository: https://github.com/safdarsidhik/murf-livekit-starter/tree/main Built with Murf Falcon — the fastest TTS API I used across this build
Part of 10 Days of Voice Agents — VoiceForBharat Edition by Murf AI
If you're building your own voice agent for an underserved use case, happy to answer questions — drop a comment.