EchoMuse

EchoMuse

Give your Amazon Echo Dot 2nd Generation a second life as an open source voice assistant satellite.

EchoMuse replaces the Alexa firmware with a lightweight Go server that streams audio to and from a backend controller of your choosing. The LED ring, microphone array, buttons, and speaker remain fully accessible via HTTP API. Pair it with Home Assistant, a local LLM pipeline, or any backend you like.

This is a significant fork of EchoGo by Binozo — the original SDK that made this hardware accessible. Substantial modifications have been made including client-side VAD, mute button handling, volume control, and audio fixes specific to the biscuit hardware.

---

Before you start

Your Echo Dot must be rooted with persistent root before EchoMuse is useful. The full rooting and setup guide — including SELinux bypass, Alexa removal, audio configuration, VAD, and wake word detection — is in SETUP.md.

The short version:

• Persistent unlock via amonet-biscuit (R0rt1z2)
• FireOS 5 (Android 5.1, API 22)
• Magisk 17.3
• Alexa voice stack disabled

---

Building

The Echo Dot runs FireOS 5 (API 22). A custom Docker build environment is required — standard Go cross-compilation won’t produce a compatible binary.

Prerequisites:

• Docker
• Go 1.24+
• GoTinyAlsa cloned to ~/GoTinyAlsa

Build the compiler image:

docker build -t echomuse-compiler compiler/

Compile:

docker run --rm \
  -e CGO_LDFLAGS="-Wl,--hash-style=both" \
  -v "$(pwd)":/sdk \
  -v ~/GoTinyAlsa:/GoTinyAlsa \
  echomuse-compiler

Output: build/server

---

Audio processing pipeline

Each 32ms period of microphone audio passes through a processing chain before VAD gating and transmission:

raw 9ch S24_3LE → beamformer (mic selection) → RNNoise NS → AGC → mono S16_LE → VAD gate

Beamformer locks to the perimeter mic with the highest onset energy ratio at voice turn start. Onset ratio (fast/slow EWMA) is robust to continuous background noise — it picks the direction that just got louder, not the loudest direction overall.

RNNoise (xiph/rnnoise v0.1) suppresses background noise. Vendored C source compiled via cgo — no external library or system dependency required.

AGC normalises level for consistent wake word and STT input. Release is frozen during silence to prevent noise floor amplification past the VAD threshold.

VAD runs on pre-NS/AGC audio to keep threshold calibrated to mic sensitivity rather than processed level.

---

VAD configuration

Tunable via environment variables on the device (overridden by controller config push):

Variable	Default	Description
VAD_CHANNEL	0	Mic channel for vad_stream endpoint (0–8)
VAD_THRESHOLD	0.015	RMS threshold 0.0–1.0 (controller pushes 0.003)
VAD_SPEECH_MS	80	Ms of speech to open the gate
VAD_SILENCE_MS	600	Ms of silence to close the gate

---

Acknowledgements

• EchoGo — Binozo
• GoTinyAlsa — Binozo
• amonet-biscuit — R0rt1z2
• EchoCLI — Dragon863
• RNNoise — Xiph.Org Foundation (BSD-3-Clause)

---

License

MIT