wilbowes/EchoMuse v2.19.0-ea.2 on GitHub

EchoMuse firmware v2.19.0-ea.2 (Early Access)

Early Access firmware: offered only by Early Access controllers, never by GA. Pair with controller 2.27.0-ea.2.

New

The Dot plays with its own voicing (#229). Measured on the Dot's own driver: a correction for its response, three compressor bands as stock has, a filter on voice replies, and a volume law of even 4 dB steps that reaches about 10 dB further at the top, where the compressors rather than the speaker take it. It runs whenever the Dot plays through its own speaker; the headphone jack keeps the plain sound. Tell us how voice and music sound.
The Echo gen 2 (radar). The firmware tells radar from the Dot 2 and opens radar's own speaker: its amp mute, the codec filter read from the unit's own /system, and a bass guard set from radar's stock tuning instead of the Dot's, which removed the woofer's bass at any normal level. Radar runs on emOS, installed by controller 2.27.0-ea.2's wizard (#687).

Bluetooth proxy scans keep running during live Sendspin music such as Spotify Connect or radio. Each scan burst is now sized from how much music is buffered, and with the proxy on the player asks Music Assistant for a 4 s buffer. A group with a proxying Echo starts, pauses and skips about 2 s later (#1004).
The first word of a command run on from the wake word is kept when the Echo detects the wake word itself: "what time is it" no longer arrives as "time is it?" (#688).
Talking over a reply cuts it on the Echo straight away, instead of a round trip later when the controller asks (#688).
Pairing needs a 7 s hold of the action button, and only asks when the Echo is not on the encrypted link. On an encrypted link the hold reaches Home Assistant as a normal long press (#841).

Dots fitted with the tsl2584tsv light sensor had no ambient light. It is now read through IIO, and only when the unit's calibration is safe to read; Amazon's other read path for it corrupts kernel memory and was behind reported reboots (#90).

Installed on a FireOS 5 Dot, emOS Dots on both kernels and an emOS radar: each found its board and light sensor by name, connected over wss, and listens with its own wake word using the streaming embedding model; each completed voice turns through Home Assistant, the Dots with the new voicing.
Radar: two units (NS6575), one provisioned by the wizard end to end; wake word, LEDs, buttons and audio checked by hand.
Sendspin, 5 min of a live source with the proxy on (emOS, 64-bit kernel): 4 s held, no underruns, adverts about 0.9/s throughout.
Run-on commands: 5 of 5 whole on a FireOS 5 Echo, 8 of 9 replayed through Home Assistant's faster-whisper.

Radar plays through the plain output chain: its speaker has not been voiced yet.
Some voice replies may sound slightly muddy with the new voicing, and it has not yet been checked for waking its own Echo at high volume.
A short command spoken over a reply still sometimes loses its start (2 of 5 "stop" captures).
An emOS Echo on the 32-bit kernel sometimes reconnects without rebooting (#816).