Native Android client for Spotify, built with Jetpack Compose
AutoMix desktop bench
Runs the phone's analysis and planning code (base/automix) on real songs and renders the
transitions to WAV files, the way the phone mixes them (B stretched with ffmpeg's atempo, pitch kept).
Prepare a folder (songs are not in the repository):
tools/automix/prepare.sh <songs folder> <work folder> # sNN.f32 per song, plus Beat This final0 references if beat_this is installed
Then, from base/:
./gradlew :automix:bench -Pargs="analyze <work folder>" # tempo, grids, downbeats; agreement with the references
./gradlew :automix:bench -Pargs="plan <work folder> s01,s02 s03,s06" # plans; beat alignment on the references
./gradlew :automix:bench -Pargs="plan <work folder> s01,s02 render" # also mix-s01-s02.wav (12 s before, 15 s after)
./gradlew :automix:bench -Pargs="matrix <work folder> long" # every ordered pair, current app's Long/Session DJ settings
./gradlew :automix:bench -Pargs="matrix <work folder> long whole" # same, preserving the whole-song exit policy
plan and matrix keep analyses in <work folder>/.automix-bench. The cache is invalidated by
analysis version, source size/modification time and model size/modification time. A warm matrix
does not decode or run ONNX again. Add fresh to matrix to force a fresh analysis. Each track's
cache/read/analysis time and each pair's planning time are printed separately, followed by counts
of BLEND/CUT/FADE and blends of at least 12 bars. These are plans, not transitions heard on a phone.
The matrix does not render audio; plan … render remains explicit.
Plan-based transitions (docs/automix/segue-design.md), from base/, with -Pthreads=<n>:
./gradlew :automix:bench -Pargs="segue <work folder> split" # tuning, validation and test folds, by artist and album, from names.txt
./gradlew :automix:bench -Pargs="segue <work folder> matrix fold=tuning" # every ordered pair: level, causes of each level left, refusals by guard, reproducibility
./gradlew :automix:bench -Pargs="segue <work folder> measure limit=100" # renders pairs in memory: peak, clipping, loudness jump, clicks, silent holes, phase against the plan's own expectation
./gradlew :automix:bench -Pargs="segue <work folder> stability limit=60" # each parameter at ±20 %: how many levels and plans change, and which variations change nothing
./gradlew :automix:bench -Pargs="segue <work folder> render s01,s02" # mix-s01-s02.wav through the same sound as the phone
names.txt has one id|title - artist.ext or id|title - artist - album.ext per song. The same
pairs are drawn on every run. Keep the test fold out of any tuning.
The bench reads the app's Beat This model (app/src/main/assets/automix/) through the JVM build of
ONNX Runtime. References: pip install beat_this (PyTorch), then beat_this <work folder>/*.wav.