Voice replies, on-device speech, message cards and 0.1.0-alpha

Gearslip now reads messages aloud and takes replies by voice. It asks
whether to reply, reads the reply back and sends it after a countdown.
Say "cancel" or "change it" to stop or redo it, or tap a quick reply.

Speech-to-text runs on the phone with whisper.cpp and a small Whisper
model, so voice works without Google services. The model ships in the
repo, so a fresh clone builds with nothing to download.

New messages show as one card at the top right, with a bar counting
down to when it closes or sends.

Other changes:
- Steering wheel buttons, knobs and D-pads drive the car screen.
- Smoother car audio: chunks wait for the car instead of dropping.
- Sharper, more responsive video.
- The Gearslip switch in Settings now stops it taking the USB
  connection.
- Apps that refuse Gearslip say so instead of loading forever.
- The phone asks for notification permission.
- The home screen stays upright.
- The car screen shows what's new after an update.
This commit is contained in:
Seb3thehacker committed 2026-10-01 00:03:53 -07:00
1 parent 1b2010f939
commit e314aa8546
46 files changed
+2290 -229

No files matched your search

+36
View File
@@ -0,0 +1,36 @@
# Speech-to-text for voice replies and the assistant: whisper.cpp, pinned to one commit and
# fetched at build time, so nothing outside this folder (cloned-repos included) is needed.
cmake_minimum_required(VERSION 3.22)
project(gearslip_speech CXX C)
set(CMAKE_CXX_STANDARD 17)
# Optimised in debug builds too. Unoptimised, Whisper runs several times slower, and a dev build
# would make voice look far worse than what ships.
foreach(lang C CXX)
set(CMAKE_${lang}_FLAGS_DEBUG "-O3 -DNDEBUG")
endforeach()
include(FetchContent)
FetchContent_Declare(
whisper
URL https://github.com/ggml-org/whisper.cpp/archive/6e4ab854f67f743900934a703d5603419384c961.tar.gz
URL_HASH SHA256=55b585aec5e1bc03e2a5cea61da97c6406fa833996cf7daeed704972273cc039
)
set(BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
set(WHISPER_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
set(WHISPER_BUILD_TESTS OFF CACHE BOOL "" FORCE)
set(WHISPER_BUILD_SERVER OFF CACHE BOOL "" FORCE)
set(GGML_OPENMP OFF CACHE BOOL "" FORCE)
set(GGML_NATIVE OFF CACHE BOOL "" FORCE)
# ARMv8.2 with half floats and dot products: every phone that runs Android 12 or later in
# practice, and what makes Whisper fast enough to answer in under a second.
set(GGML_CPU_ARM_ARCH "armv8.2-a+dotprod+fp16" CACHE STRING "" FORCE)
FetchContent_MakeAvailable(whisper)
add_library(gearslip_speech SHARED speech_jni.cpp)
target_compile_options(gearslip_speech PRIVATE -O3)
find_library(LOG_LIB log)
target_link_libraries(gearslip_speech whisper android ${LOG_LIB})
target_link_options(gearslip_speech PRIVATE -Wl,--gc-sections -Wl,-z,max-page-size=16384)