Deepgram vs Hume AI

Developer Builds Multilingual AI Voice Agent 'Sadie'

A developer, Ayush Shinde, built Sadie — a multilingual AI voice agent — during the "10 Days of AI Voice Agents — #VoiceForBharat Edition" challenge. Over ten days he implemented real-time voice conversations, Hindi-English code-mixing support, persistent (permissioned) memory, tool integration, outbound phone calling via SIP, human escalation via Discord webhooks, an analytics dashboard, and multi-agent specialist handoffs (grammar, maths, full-stack). The system integrates LiveKit for real-time transport, Deepgram for STT, Google Gemini as the LLM brain, and Murf Falcon for TTS, with SQLite for memory and Next.js/Chart.js for analytics. The post documents architecture, implementation details, encountered bugs (telephony SIP encryption, logging issues, import paths), and instructions to run the public repository and demo.

• Ayush Shinde built an AI voice agent named Sadie during the "10 Days of AI Voice Agents — #VoiceForBharat Edition" challenge.
• Sadie supports real-time voice, Hindi-English code-mixed conversations, permissioned persistent memory (SQLite), external tool integration, outbound phone calls, human escalation, analytics, and specialist agent handoffs.
• Core technical stack includes LiveKit (real-time audio), Deepgram (Speech-to-Text), Google Gemini (LLM), Murf Falcon (TTS), SQLite, Next.js, and Chart.js.

10-Day AI Voice Agent Stops Scam Calls

A developer built 'Suraksha Saathi', an AI voice agent prototype to help families detect and respond to scam calls in 10 days. The agent uses Deepgram for speech recognition, a Google Gemini model for reasoning, Murf AI's Falcon TTS for fast speech output, and LiveKit for real-time connectivity. Features include Hindi support via Devanagari-only output, a consent-based memory to remember callers, a panic button that alerts a human fraud team via Discord, real-time scam tracking dashboard, and mid-call voice-engine handoffs to escalate to an investigator persona. The author published the post on DEV and shared source code on GitHub.

• Author built an AI voice agent named Suraksha Saathi in 10 days.
• Architecture: Deepgram (ASR) → Google Gemini (LLM) → Murf AI Falcon (TTS) → LiveKit (real-time audio).
• Agent enforces Hindi output in Devanagari, implements consent-based memory, and a panic button that notifies a human fraud team via Discord.

Voice AI Assistant for Indian Farmers Built in 10 Days

A developer built "Farm & Field," a voice AI assistant for Indian farmers during the 10 Days of Voice Agents — VoiceForBharat Edition by Murf AI. The system integrates Speech-to-Text, an LLM, Text-to-Speech, and real-time transport (Deepgram, Google Gemini, Murf Falcon, LiveKit) with SQLite-backed memory, weather and market-price tools, human escalation, a specialist crop agent, outbound calling via LiveKit SIP, and call analytics. The agent supports Malayalam, English, and Manglish, requests consent before storing memory, and hands off to humans or specialists when appropriate. The project code is available on GitHub.

• Author built Farm & Field during the 10 Days of Voice Agents — VoiceForBharat Edition by Murf AI.
• Core stack: Deepgram (STT), Google (Gemini) for LLM, Murf Falcon (TTS), and LiveKit (real-time transport).
• Features include multilingual Malayalam/English/Manglish support, memory-with-consent, weather and market-price tools, human escalation, and a Crop Problem Specialist agent.

Evaluating Google’s multi-speaker TTS: A case study in why private evaluations matter

Hume AI published a new blog article titled 'Evaluating Google’s multi-speaker TTS: A case study in why private evaluations matter'.