Repository navigation
0.3.0 — reliability, datasets, recording, queues
The release where the claims got measured and enforced.
Reliability
SoundRsError.code:io,format,device,stalled,unsupported,usage- no panics on corrupt input: 1000 mutated fixtures per run in
cargo test - deterministic decode pinned by PCM hashes in CI (WAV/FLAC/AAC exact, MP3 by length + RMS)
Player.underrunsandPlayer.latencyexpose the stream's health- player queue handover waits for the device to drain;
feed()times out instead of hanging
Datasets and ML
load(path, sample_rate=16000, mono=True)and the same options onload_manyload_bytesfor data that is already in memory- parallel decode still uses every core with the GIL released
Recording
Recorder+list_input_devices():record(seconds),start()/stop(), context manager
Playback
Player.enqueuechains tracks across different rates and layoutsbegin_feed/feed/end_feedpush samples from Python with backpressure
Benchmarks
- subprocess-isolated harness with warmup, median/p95/p99, bootstrap CI, CPU time, peak RSS,
JSON results, baseline diffing and a markdown report (bench/run.py) - new scenarios: real-time factor, underruns, thread scaling, cold/warm, 200-cycle soak
Packaging
- wheels now use the stable ABI (
abi3-py310): six files instead of sixteen — Linux x86_64
and aarch64, Windows amd64, macOS arm64 and Intel, each for Python 3.10+
Full build report, including the review findings that were fixed: docs/reports/0.3.0-report.md
Docs: https://neksand.github.io/sound_rs/