EchoMuse firmware v2.19.0-ea.1 (Early Access)
Early Access firmware: offered only by Early Access controllers, never by GA. Pair with controller 2.27.0-ea.1.
New
- Barge-in over music (#975). The speaker loopback the echo canceller learns from carried the music's highs folded into the speech band, where the microphones never hear them, so the canceller chased signal that was not there. The firmware now rebuilds that reference from its own output. Barge-in over loud, bass-heavy music at 70% volume went from 1 in 5 to 4 in 5. Echo paths saved by earlier firmware are discarded once, since they were learnt against the old reference.
- Cheaper wake word scoring (#893, #894). With the streaming embedding model installed by controller 2.27.0-ea.1, the Echo computes only the audio that is new on each step: about half the scorer's CPU on a FireOS 5 Echo, with the same scores. Without that file, or if it fails its check at load, the Echo uses the stock model.
Tested
- Echo cancellation of pink noise through the firmware went from 1 dB to 25 dB on FireOS 5, 22 dB on emOS (32-bit kernel) and 17 to 22 dB on emOS (64-bit kernel).
- The streaming model gives the stock model's scores on the host and on a FireOS 5 Echo through the probe, and drops the scorer from 41% to 23% of one core.
Known issues
- It has not yet been confirmed from an installed Echo's log that the streaming model, rather than the stock fallback, is the one in use.
- An emOS Echo on the 32-bit kernel sometimes reconnects without rebooting (#816).