Share lossless voice prefixes across companion announcements

This commit is contained in:
mikecarper committed 2026-10-03 19:11:15 -07:00
1 parent 91a9190a71
commit 22360f4392
18 files changed
+2284 -2018

No files matched your search

+18
View File
@@ -77,6 +77,24 @@ Power-on and wake wiring remain board-specific. Speech is nonblocking and uses
the existing nRF52 PWM player; each buzzer still needs a physical listening
test for intelligibility. All spoken button, alert-mode and GPS recordings use
Microsoft Zira Desktop with the selected compressed Companion voice profile.
GPS on/off use three stored audio spans: one shared encoded GPS prefix and two
phrase tails. Advert queued/failed also share an identical encoded prefix;
Sound off/Sound only share one only on `HAS_DRV2605` boards. The tails are not
independently decodable words. Playback treats each prefix and tail as one
continuous compressed stream, without resetting its codec, bit position,
interpolation or DMA at the seam. The approved full-phrase PCM and PWM output
remain bit-for-bit identical, including pronunciation and pauses.
The IMA audio payload is 3,881 bytes smaller on ordinary voice boards, or
4,971 bytes smaller on vibration-equipped boards, before the small segmented
descriptor/decoder overhead. Tiny shared silence prefixes are not split.
A representative Cortex-M4 GCC 14.2 size-optimized/LTO player-and-catalog
build saves 3,732 or 4,792 bytes respectively after that overhead; final
firmware savings can vary with toolchain and surrounding code.
The existing generators still synthesize the two complete GPS phrases and
button confirmations, then share byte-identical encoded prefixes; they do not
try to identify word boundaries or splice independently spoken words. Windows
Zira synthesis still requires Windows' local `System.Speech` tooling.
R1 Neo's buzzer pin aliases its optional QSPI clock, so voice is
not default-enabled when `QSPIFLASH` is selected. ThinkNode M4 is deliberately
excluded: its former