This is an ongoing repository with test scripts and experiments for the Amazon Echo Dot 2 (biscuit).

Live demo: Watch on YouTube Shorts

Getting Started

To run these scripts, you must first unblock the bootloader and gain root access on your Echo Dot 2.

1. Unlock & Root: Follow the firmware unlocking guide available at XDA Forums.
2. Required Scripts: You will need to flash the exploit script downloadable from this specific post.

Note for Mac Users: Even though the original guide focuses heavily on Linux, the scripts provided on the XDA forum also work on Mac! I was able to follow along and execute the exploit steps successfully on macOS.

Contents

• test-scripts/ - Contains shell scripts to test device I/O (buttons, LEDs, etc.) natively on the Echo Dot 2.
  • connect-to-wifi.sh: Connects the Echo Dot to a Wi-Fi network by injecting credentials into wpa_supplicant.conf via ADB and requesting a DHCP lease via dhcpcd.
• debloat/ - Contains soft-debloat.sh, a script to dynamically disable/hide unnecessary Amazon packages, rename bloat binaries, replace bootanimation with a no-op stub, and patch /system/etc/audio_init.sh to permanently disable the infinite spinning cyan LED ring via hardware sysfs (is31fl3236).
• custom-voice-assistant/ - Suite containing the custom assistant (echo-dot-assistant) and dedicated multi-process Pryon/Alexa wake-word service (echo-dot-wakeword), managed by a unified parent Makefile.
• reverse-engineer-speech-interaction/ - Decompiled sources, native library JNI probes, and reverse-engineering logs for Fire OS's native speech and wake-word stack (libwakewordserver_jni.so + libpryon.so).

Proprietary Amazon files

The repository does not include Amazon's proprietary APKs or native .so libraries. They are referenced by name in the notes only. These files are copyrighted and must not be shared or redistributed. To run the wake-word experiments locally, pull the required files from your own rooted Echo Dot or extract them from a firmware image you obtained yourself.

The tested firmware was update-kindle-csm_biscuit-272.6.8.0_user_680767620.bin, available through the Echo Dot 2 rooting thread on XDA. From reverse-engineer-speech-interaction/, use ./pull-and-decompile.sh --find to locate candidate files on a connected Echo, ./pull-and-decompile.sh --pull-native to pull them into the ignored local-native-libs/ directory, or ./pull-and-decompile.sh --extract-apk SpeechInteractionManager.apk after obtaining the APK yourself.

Echo Dot Custom Voice Assistant Stack

The overall architecture replaces Amazon's cloud stack with a fully local, privacy-focused pipeline running on the Echo Dot 2 (API 22, Android 5.1, armeabi-v7a):

1. Wake-Word Engine (echo-dot-wakeword):
   • Installed as a system app (/system/priv-app/CustomWakeWord) with CAPTURE_AUDIO_HOTWORD permissions.
   • Interfaces directly with native libpryon.so / libwakewordserver_jni.so to run low-power local wake word detection ("Alexa").
2. Speech-to-Text (echo-dot-assistant):
   • Uses native Sherpa-ONNX (libsherpa-onnx-jni.so linked against ONNX Runtime 1.14.0) for fast, offline streaming ASR.
3. Local LLM Engine (llama-server):
   • Uses an app-managed, background llama-server running MobileLLM-125M-Q4_K_M.gguf.
   • Uses prompt caching (cache_prompt=true) to reduce warm query latency from 17.3s (cold llama-cli) down to **2.3s**.
   • Outputs JSON-constrained function calls (e.g., LIGHT or VOLUME actions).

Development & Automated Testing

A parent Makefile in custom-voice-assistant/ simplifies building, deploying, and testing all components:

cd custom-voice-assistant

# Build and install both assistant and wake-word APKs
make deploy

# Run automated device test suite
make test_all     # Runs wake-word, STT, and LLM tests end-to-end
make test_stt     # Tests Sherpa-ONNX speech recognition
make test_llm     # Tests llama-server inference & command parsing
make test_wake    # Tests native Pryon wake-word service

Manual ADB Pipeline Testing

You can also trigger pipeline components programmatically via ADB intents:

Testing STT Pipeline:

adb shell am force-stop com.custom.assistant
adb shell logcat -c
adb shell am startservice -n com.custom.assistant/.AudioService --es TEST_MODE STT
adb shell logcat -v time -s CustomAssistant AndroidRuntime

Testing LLM Pipeline:

adb shell am force-stop com.custom.assistant
adb shell logcat -c
adb shell 'am startservice -n com.custom.assistant/.AudioService --es TEST_MODE LLM --es TEST_TEXT "turn on the light"'
adb shell logcat -v time -s CustomAssistant AndroidRuntime

System Requirements & Binary Dependencies

1. llama-server / llama-cli: Compiled for armeabi-v7a and stored at /cache/llama-server (or /cache/llama-cli).
2. LLM Model: MobileLLM-125M-Q4_K_M.gguf stored at /cache/MobileLLM.gguf.
3. STT dependencies: Sherpa-ONNX, ONNX Runtime and their Java/native build artifacts are third-party dependencies and are not included in this repository. Obtain them from their upstream projects and place them in the paths expected by echo-dot-assistant/Makefile. They remain under their respective upstream licenses.

Copyright notice

This repository contains original project code and documentation, plus separately licensed third-party open-source components where applicable. It does not distribute Amazon firmware, Amazon APKs, Amazon native libraries or user data. The extraction scripts only operate on files obtained locally by the user from their own device or firmware copy. Amazon, Alexa and related names are trademarks of Amazon.com, Inc. This project is unofficial and is not affiliated with or endorsed by Amazon. If a copyright holder believes that any material has been included improperly, please contact the maintainer so it can be reviewed and removed.