HearoPilot is an open-source Android app that provides real-time audio transcription and AI-generated meeting insights — summaries, action items, and key points — entirely on your phone. It uses Sherpa-ONNX for speech-to-text and llama.cpp for LLM inference, with zero data sent to any server.

Does HearoPilot send my meeting audio to the cloud?

No. HearoPilot is 100% offline. All transcription and AI processing happens on-device using local models. No audio, transcripts, or insights ever leave your phone. This makes it suitable for confidential meetings, HIPAA-adjacent conversations, and enterprise environments with strict data policies.

What AI models does HearoPilot use?

For speech-to-text, it uses NVIDIA's NeMo Parakeet TDT 0.6B (Int8 quantized, ~670MB) via Sherpa-ONNX. For LLM insights, it runs Google's Gemma 3 1B via llama.cpp — available in Q8_0 (~1GB) for higher quality or IQ4_NL (~650MB) for mid-range devices. Total storage needed is about 1.7GB.

What recording modes does HearoPilot support?

Four modes: Simple Listening (lightweight transcription), Short Meeting (standups and 1:1s), Long Meeting (extended sessions with lower-frequency insights), and Real-Time Translation (live translation across 25 supported languages).

What devices can run HearoPilot?

Any Android device running Android 11 or later (API 30+) with approximately 1.7GB of free storage for models. The app automatically recommends the best LLM variant based on your device's RAM and Android version. A physical device is strongly recommended over an emulator.

How does HearoPilot compare to Otter.ai or Fathom?

Otter and Fathom are cloud-based — they join meetings as visible bots and process audio on remote servers. HearoPilot runs entirely on your phone with no cloud dependency, no visible bot, and no data leaving the device. The tradeoff is that cloud tools offer team collaboration features and multi-platform support, while HearoPilot prioritizes privacy and offline operation.

HearoPilot: On-Device AI Meeting Assistant for Android

By Prahlad Menon Published 2026-04-11 4 min read

A few weeks ago I wrote about the AI meeting copilot landscape — how tools like OpenOats and Meetily are proving that real-time transcription and summarization can run locally, while cloud incumbents like Otter, Fathom, and Fireflies still route your conversations through someone else’s servers. Every tool in that roundup was either desktop-native or cloud-dependent. The mobile gap was glaring.

HearoPilot fills it. Built by de.ai (Decentralized AI), it’s an Android app that runs real-time speech-to-text and LLM-powered meeting insights entirely on your phone. No cloud calls. No meeting bots joining your Zoom. No data leaving the device. Period.

It’s available now on the Google Play Store, and the full source is on GitHub under the Apache 2.0 license.

What It Actually Does

HearoPilot offers four recording modes designed around different use cases:

Simple Listening — quick capture and transcription
Short Meeting — optimized for standups and 1:1s
Long Meeting — handles extended sessions with smart chunking
Real-Time Translation — live translation across languages as people speak

After recording, the on-device LLM generates structured insights: summaries, key points, action items — the same outputs you’d expect from Otter or Fathom, except nothing touches a server.

The app supports 25 UI languages with localized LLM system prompts, meaning the AI doesn’t just translate — it thinks in the target language’s context. That’s a level of internationalization you rarely see in meeting tools, cloud or otherwise.

The Technical Stack Is Serious

What makes HearoPilot genuinely impressive isn’t the feature list — it’s the engineering underneath.

Speech-to-text runs on Sherpa-ONNX with NVIDIA’s NeMo Parakeet TDT 0.6B model, quantized to Int8 (~670MB). This is a production-grade ASR pipeline running on a phone, not a toy demo.

LLM inference uses llama.cpp running Google’s Gemma 3 1B — available in Q8_0 (~1GB) or the more aggressive IQ4_NL quantization (~650MB) for devices with less headroom. The total model footprint is roughly 1.7GB of storage.

The optimizations are where the real craft shows:

VAD (Voice Activity Detection) pipeline filters silence and noise before it hits the STT model, saving compute cycles and improving accuracy
KV cache reuse across LLM inference calls avoids redundant computation — critical when you’re generating insights from a long transcript
Thermal throttling awareness — the app monitors device thermals and adapts inference speed to prevent your phone from becoming a space heater
Adaptive memory management keeps the app stable on devices with varying RAM availability

The codebase itself is clean: Kotlin with Jetpack Compose, Hilt for dependency injection, Room for local persistence, and Material Design 3. It’s textbook Clean Architecture, which matters because it’s open source — people will actually read and fork this code (24 forks already).

Why This Matters

The cloud meeting AI market has a fundamental trust problem. When Otter transcribes your board meeting or Fireflies records your therapy-adjacent 1:1 with a direct report, that audio hits external servers. You’re trusting a startup’s security posture with your most sensitive professional conversations.

Desktop local-first tools like OpenOats and Meetily solved this for laptop users. But meetings happen everywhere — in coffee shops, on walks, in cars. Your phone is the device that’s always with you, and until now, there was no serious option for on-device meeting AI on mobile.

HearoPilot changes that. Running a 0.6B parameter STT model and a 1B parameter LLM on a phone — with thermal management, VAD, and KV cache optimization — is a legitimate engineering achievement. This isn’t a research project; it’s a usable app on the Play Store.

The Bottom Line

Requirements: Android 11+ (API 30), ~1.7GB storage for models, and a reasonably modern chipset.

License: Apache 2.0 — use it, fork it, build on it.

Links: GitHub · Google Play Store

If you’ve been waiting for a privacy-respecting meeting assistant that lives in your pocket, this is the one to watch. The local-first AI meeting space just went mobile.