Open Source Voice Assistant for Android in 2026

October 6, 2026 · guide · 7 minutes read

Proprietary voice assistants are convenient until you ask where your conversations go and why you can't add one simple action. Open source solves both problems. In 2026, there's finally a voice assistant for Android that's fast, private, and actually useful.

Why Open Source Matters for Voice Assistants

Your voice is the most personal interface. Every command, every question, every casual "set a timer for 10 minutes" goes through a black box. With proprietary assistants:

You don't know where your recordings are stored or who can access them
You can't add custom actions without waiting for a corporate roadmap
You're locked into an ecosystem — change your phone, lose your assistant
Privacy policies change, telemetry gets added, and you can't opt out

Open source voice assistants let you audit the code, control your data, and extend functionality yourself. MIT license means you can fork it, modify it, ship it. No tracking, no analytics, no "improving the product".

Amalia: Speed and Privacy by Design

Amalia is an open source voice assistant for Android built with two non-negotiable principles:

Speed: Full voice cycle (end of speech → first sound of response) in ~0.87 seconds. That's faster than Google Assistant.
Privacy: Conversation history stored only on device. Zero analytics, zero telemetry. Your API keys — your control.

What Makes It Fast

Most assistants feel slow because they wait for everything to finish before starting the next step. Amalia pipelines the stages:

Voice detection: Silero VAD runs offline on device. ~200ms to detect end of speech.
Speech recognition: Groq Whisper v3 Turbo. ~220ms to transcribe.
Brain: Qwen3 27B on Groq LPU. ~150ms to first token.
Voice synthesis: Fish Audio drama-3 starts streaming before the model finishes thinking.

Total: 0.87 seconds. The technical breakdown is in this latency analysis.

What Makes It Private

Amalia's privacy model is simple and auditable:

Voice activity detection (VAD) runs entirely offline using Silero
Speech recognition and LLM inference happen on Groq (you control the API key)
TTS synthesis happens on Fish Audio (again, your API key)
Conversation history is stored locally in DataStore, encrypted, never leaves the device
No analytics. No crash reporting. No "sign in to sync"

The trade-off is honest: for sub-second latency, you need internet. But the code is open — you can verify exactly what's sent and to where. Full privacy analysis: where your conversations go.

What Amalia Can Do

Unlike command-only assistants, Amalia is conversational AI with system access:

Phone control: Wi-Fi, Bluetooth, brightness, volume, flashlight, alarms, timers
Battery status: Ask how much charge is left
Natural language: No rigid command syntax — speak naturally, it understands
Personality: Amalia has character. She's direct, sometimes sharp, never apologizes for being an AI

Full command list: voice commands reference.

How It Compares to Alternatives

vs Google Assistant

Speed: Amalia is faster (~0.87s vs 1.5-2s)
Privacy: Google records everything. Amalia keeps history on device
Customization: Amalia is MIT-licensed — fork and modify. Google is a black box
Trade-off: Google has deeper system integration and smart home ecosystems

vs Other Open Source Assistants

Mycroft/OpenVoiceOS: Platform for smart home devices, not a phone app
Rhasspy: Fully offline but command-based, no conversational AI
Dicio: Lightweight Android assistant, but limited dialogue capabilities

Detailed comparison: open source assistants overview.

How to Get Started

Amalia runs on Android 8.0+. You'll need:

The app: Download APK (10 MB)
Groq API key: Free tier available, covers typical usage
Fish Audio key: Also free tier

Keys are entered in the app settings — they're stored locally and never leave your device except for API requests. If you leave the fields empty, the app uses the bundled keys from the build.

No Google Account Required

Unlike most Android assistants, Amalia doesn't require Google services. Sideload the APK, add your keys, and you're done. Works on de-Googled phones.

For Developers: MIT License, Clean Architecture

The codebase is Kotlin 2.2 + Jetpack Compose + Material 3. No DI frameworks, no bloat — ServiceLocator, Coroutines, and a clean state machine for the voice pipeline.

Repository: kurumi-mProject/amalia
License: MIT — fork it, ship it, just keep attribution
Builds: GitHub Actions CI produces APKs automatically

Want to build your own? How to build a voice assistant walks through the full stack.

The Trade-Offs You Should Know

Amalia is fast and private, but it's not magic:

Requires internet: STT, LLM, and TTS happen in the cloud. Offline-first users should look at Rhasspy or Dicio.
API costs: Free tiers cover normal use, but heavy usage costs money. Cost breakdown here.
Limited smart home: No native Home Assistant integration yet. Phone control only.
No wake word: Tap and hold to activate. Wake word detection is on the roadmap.

These are design choices, not bugs. Amalia optimizes for speed and conversational quality, which requires cloud inference. The code is open — if you need offline, you can swap Groq for whisper.cpp and a local LLM.

Why This Matters in 2026

Google Assistant is being replaced by Gemini. Alexa is full of ads. Siri still can't set two alarms. The trend is clear: voice assistants are becoming content recommendation engines and shopping portals.

Amalia exists because voice should be a tool, not a product. You own the code. You control the data. You decide what gets added.

If that sounds like what you need, download the APK and see how fast an open source assistant can be.