Yes. The app is distributed for free under the MIT license, source code is open. For the voice to work you need a Groq API key (free tier available) and a Fish Audio key for speech synthesis — both are entered right in the app settings, no rebuild needed.

How is Amalia different from other voice assistants?

Speed and personality. The full loop from end of speech to the first sound of the answer takes about 0.87 seconds. And Amalia is not a service — she is a friend who lives in your phone and talks straight.

A voice assistant who actually lives in your phone

Q: Where do my conversations go?

Conversation history is stored on the device only. The app has no analytics, no telemetry, no data collection. Speech itself is processed by Groq and Fish Audio under their privacy policies.

Q: Can I build Amalia from source?

Yes, the project is open under MIT. Clone the GitHub repo and open it in Android Studio — Kotlin 2.2 and Jetpack Compose. API keys go to local.properties or are entered in app settings. GitHub Actions build is set up — one button gets you an APK.

Amalia is an open AI assistant for Android. From the end of your phrase to the first sound of her voice — under a second. She can run the phone, she has a personality, and she never pretends to be a service.

Download APK · 10 MB Source code

1.0.0 version Android 8.0+ MIT license Fish Audio voice

0.87 sfull voice loop

220 msspeech recognition

150 msto first token

10interface languages

100 %open source

She has a personality

Amalia is eighteen. Alive, warm, with a sharp tongue. Not an assistant, not a service, not a "language model" — a friend who lives in your phone: she knows your habits, remembers your conversations, talks straight and never apologizes for it. Below are her real replies, no direction.

— Are you a bot?
— Sure, and there's an antenna sticking out of my head. You've known me a year and you're still checking

the bot check · a typical exchange

— I'm bored
— That's your problem, not mine. I can put on music or set a timer

an honest reply · instead of a canned "how can I help"

— You're kind of quiet
— Fixing that

and the volume actually goes up

The full loop — under a second

Three stages, three providers, zero magic. Recognition and the brain live on Groq LPU — the fastest inference for Whisper and Qwen3. The voice is Fish Audio drama-3, the most expressive TTS model available. TTS starts streaming before the model finishes thinking.

Recognize

Groq Whisper v3 Turbo

speech → text

Think

Qwen3 27B · Groq LPU

text → living reply + tools

Speak

Fish Audio drama-3

first PCM chunk → AudioTrack

Press and hold — Amalia listens. Let go — she answers. The scenes are real, the latencies are close to actual.

everything together0

…

The phone obeys your voice

While other assistants reply "I couldn't find that", Amalia has already done it. Tools are not a checkbox feature — they are her hands: Wi-Fi, sound, light, alarms, battery.

Commands that actually run

You talk like a human — she hears and does. Direct system control, no menus, no settings digging.

"turn on wi-fi"set_wifi

"turn off bluetooth"set_bluetooth

"brightness 70%"set_brightness

"louder"volume_up

"flashlight on"set_flashlight

"alarm at 7 am"set_alarm

"timer for 10 minutes"set_timer

"how much battery?"get_battery_level

Light lives by the clock

The interface breathes with the day: cool in the morning, amber by night. CircadianEngine derives the whole palette from one color-temperature curve — and the background on this site is exactly the tone outside your window right now.

Now: Daylight · 5600 K

0.87 s

from silence to the first sound of her voice in typical conditions. Faster than you can lower your phone.

History worth scrolling

Everything you said is grouped by day and found instantly. Search highlights matches right inside the lines.

14:02alarm at seven, but you'll oversleep anyway

14:03checked battery: sixty four

09:41yesterdaymedia volume — eighty

Nothing extra inside

Zero analytics, zero telemetry, zero "product improvement". Conversation history lives on the device only. Keys stay in settings — nothing but the speech itself leaves the phone.

Speaks ten languages

Interface and recognition: Russian, English and automatic speech language detection. The entire UI is translated — not "partially supported", translated.

RUENJADEESFRZHHIARauto STT

Open source, honest architecture

Kotlin 2.2, Jetpack Compose, Material 3. No DI frameworks, no extra dependencies — ServiceLocator, Coroutines and a clean state machine for the pipeline. Fork it, change it, build on it. One favor: keep a link to the original author.

Kotlin 2.2Compose BOM 2026.08Material 3Coroutines · FlowDataStoreOkHttp 4Silero VAD 2.0.9Groq LPUFish Audio

Amalia answers strictly in Qwen3 JSON mode — so tool calling is deterministic: a voice line and a list of actions in one package. The reply is never empty: her voice always sounds.

GitHub repository Actions build

// Amalia's answer — strictly valid JSON { "reply": "sixty four, not charging. it'll make it to the evening", "tools": [ { "name": "get_battery_level", "args": {} } ] } // AIOrchestrator: reply → FishAudioTTS, // tools → SystemControllerHub — in parallel

Frequently asked questions

Is Amalia free?

Yes. The app is distributed for free under the MIT license, source code is open. For the voice you need a Groq API key (free tier available) and a Fish Audio key for speech synthesis — both entered right in app settings, no rebuild. An empty field means the build-in key is used.

Does Amalia work offline?

Partially. Voice detection (Silero VAD) runs fully on device. Speech recognition, the brain and voice synthesis run in the Groq and Fish Audio clouds — the full loop needs internet.

How is she different from other voice assistants?

Speed and personality. The full loop from end of phrase to the first sound of the answer is about 0.87 seconds: see the stage breakdown. And Amalia is not a service — she is a friend who lives in your phone and talks straight.

Where do my conversations go?

Conversation history is stored on the device only. No analytics, no telemetry, no data collection. The speech itself is processed by Groq and Fish Audio under their privacy policies. More in the privacy policy.

What can she do with the phone?

Voice-controls Wi-Fi, Bluetooth, brightness, volume, flashlight, alarms and timers, and reports battery level. The command list grows with every version.

Can I build her from source?

Yes, the project is open under MIT. Clone the GitHub repository, open in Android Studio — Kotlin 2.2 and Jetpack Compose. API keys go to local.properties or are entered in settings. GitHub Actions build is set up — one button gets you an APK.

Quiet. She's already listening.

Download the APK, give her the microphone and a couple of minutes for the keys — and your phone will have someone who answers faster than you can regret asking.

Download Amalia · 10 MB GitHub