Audio refactor

README.md

@@ -66,7 +66,7 @@ shape: (5_271_939, 3)
 
 ```pycon
 >>> x['audio']
-WSAudio(audio_reader=AudioReader(src=<class '_io.BytesIO'>, sample_rate=None), tstart=614.46246, tend=627.3976)
+WSAudioSegment(episode=WSAudioEpisode(src=<class 'pyarrow.lib.BufferReader'>, sample_rate=None), tstart=614.46246, tend=627.3976)
 
 ```
 

requirements.txt

@@ -4,5 +4,4 @@ numpy
 polars>=1.36.1
 pyarrow>=20
 torch
-torchaudio
 # torchcodec – optional, it causes serious performance regressions

scripts/test_audio_backends.sh

@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+# Run the full test suite in isolated environments, one per decoder backend.
+#
+# Usage:
+#   ./scripts/test_audio_backends.sh
+#
+# Each environment installs only one decoder backend so we verify that the
+# interfaces work correctly regardless of which backend is present.
+
+TEST_COMMAND=${1:-"python -m tests"}
+
+parallel --tag --lb \
+    "uv run --isolated --with {} $TEST_COMMAND" \
+    ::: humecodec "torchaudio<2.9" torchcodec

tests.py

@@ -2,14 +2,16 @@
 import unittest
 
 import wsds
-from wsds import ws_dataset, ws_shard, ws_sink
+from wsds import ws_dataset, ws_shard, ws_sink, ws_audio, audio_codec
 
 
 def load_tests(loader, tests, ignore):
     tests.addTests(doctest.DocTestSuite(wsds))
     tests.addTests(doctest.DocTestSuite(ws_dataset))
     tests.addTests(doctest.DocTestSuite(ws_shard))
     tests.addTests(doctest.DocTestSuite(ws_sink))
+    tests.addTests(doctest.DocTestSuite(ws_audio))
+    tests.addTests(doctest.DocTestSuite(audio_codec))
     tests.addTests(doctest.DocFileSuite("README.md"))
     return tests