[ingoau]

Why

AutoMix across sample rates still left a short gap now and then. The cause is ExoPlayer's DefaultAudioSink, not the mixer: it reuses its AudioTrack only when the output format is unchanged (canReuseAudioOutput). At a 44.1 ↔ 48 kHz join it played A out to silence, released the track and built a new one. That leaves a gap between tracks, and in the middle of A's audio when the join is a mix.

What

MixProcessor now always outputs at one rate, so every join is an ordinary gapless join and the sink keeps its track.

• Output rate (OutputRate.kt): the first stream's rate after a reset when it is 44.1 or 48 kHz, else the device's (PROPERTY_OUTPUT_SAMPLE_RATE). A library at one rate is never resampled and stays bit-exact. There's no extra CPU and no double resampling over 44.1 kHz Bluetooth.
• Resampler (crates/hocket-android/src/resampler.rs, JNI): a Kaiser-windowed sinc with exact taps per output phase (160 rows for 44.1 ↔ 48 kHz, ~80 dB stopband). Output frame j is input frame j × in / out exactly. It holds back only the kernel's ~1.5 ms lookahead, not a 1024-frame chunk. It runs after the mixer, so the mixer's sample-exact item accounting is unchanged.
• Seams:
  • A gapless join at the same format keeps going through the same resampler, so there's no seam inside continuous audio.
  • When the next stream is at another format (configured while A plays), A's end of stream lets out A's last frames. When the mix follows, it uses the armed tail's after (A's real continuation) as lookahead. B's resampler is primed with the tail's lead, so the switch from A to the tail is as clean as at one rate.
• Tail render: automix_prepare_tail decodes from 60 ms before the tail start. That gives the tail's first frames filter history; before, they were resampled from silence. It returns lead and after along with the tail.

Testing

• cargo test -p hocket-android (26 tests). New resampler tests cover:
  • exact output lengths
  • sine accuracy and no delay
  • alias rejection
  • chunking invariance
  • finish with lookahead and prime with history both matching an uncut stream
  • bounded memory
  • lead/after alignment in prepare_tail
• :playback:testDebugUnitTest: new OutputRateTest (11 tests) covers:
  • latching
  • passthrough
  • a same-format join continuing
  • a seek restarting
  • exact finish with lookahead
  • priming
  • the final track coming out whole at the end of the playlist
  • a lead still priming a resampler that hasn't had input yet
  • in MixProcessor: a fixed output format across items, end of stream holding isEnded until A's last frames are out, and the resampler staying open at a same-format end of stream
• Not tried on a device or emulator yet.

Known limits

• A channel-count change (mono ↔ stereo) still rebuilds the AudioTrack.

🤖 Generated with Claude Code