add require_audio

Author: eustlb

src/transformers/audio_utils.py

@@ -196,7 +196,7 @@ def _resolve_audio_source(audio: str, timeout: float | None = None) -> "str | by
 def _load_audio_soundfile(source: "str | bytes", sampling_rate: int) -> np.ndarray:
     """Decode with soundfile and resample with soxr: the same steps as `librosa.load`, so the output is identical
     (downmix by averaging, `soxr_hq` resampling, padding/trimming to `ceil(num_samples * sampling_rate / src_rate)`)."""
-    requires_backends(load_audio, ["soundfile", "soxr"])
+    requires_backends(load_audio, ["audio"])
     with sf.SoundFile(BytesIO(source) if isinstance(source, bytes) else source) as f:
         src_sampling_rate = f.samplerate
         waveform = f.read(dtype="float32", always_2d=True).T

src/transformers/testing_utils.py

@@ -80,6 +80,7 @@
     is_apex_available,
     is_apollo_torch_available,
     is_aqlm_available,
+    is_audio_available,
     is_auto_round_available,
     is_av_available,
     is_bitsandbytes_available,
@@ -153,7 +154,6 @@
     is_seqio_available,
     is_serve_available,
     is_soundfile_available,
-    is_soxr_available,
     is_spacy_available,
     is_speech_available,
     is_spqr_available,
@@ -887,6 +887,14 @@ def require_torchcodec(test_case):
     return unittest.skipUnless(is_torchcodec_available(), "test requires Torchcodec")(test_case)
 
 
+def require_audio(test_case):
+    """
+    Decorator marking a test that requires the audio dependencies (soundfile and soxr, installed with
+    `transformers[audio]`). These tests are skipped when they aren't installed.
+    """
+    return unittest.skipUnless(is_audio_available(), "test requires audio")(test_case)
+
+
 def require_torchaudio(test_case):
     """
     Decorator marking a test that requires torchaudio. These tests are skipped when torchaudio isn't installed.
@@ -1828,13 +1836,6 @@ def require_soundfile(test_case):
     return unittest.skipUnless(is_soundfile_available(), "test requires soundfile")(test_case)
 
 
-def require_soxr(test_case):
-    """
-    Decorator marking a test that requires soxr
-    """
-    return unittest.skipUnless(is_soxr_available(), "test requires soxr")(test_case)
-
-
 def require_multipart(test_case):
     """
     Decorator marking a test that requires python-multipart

src/transformers/utils/__init__.py

@@ -119,6 +119,7 @@
     is_apex_available,
     is_apollo_torch_available,
     is_aqlm_available,
+    is_audio_available,
     is_auto_round_available,
     is_av_available,
     is_bitsandbytes_available,

src/transformers/utils/import_utils.py

@@ -1662,6 +1662,13 @@ def is_soxr_available() -> bool:
     return _is_package_available("soxr")[0]
 
 
+@lru_cache
+@_make_compile_constant
+def is_audio_available() -> bool:
+    """The audio core's dependencies, installed with `transformers[audio]`: soundfile to decode, soxr to resample."""
+    return is_soundfile_available() and is_soxr_available()
+
+
 @lru_cache
 @_make_compile_constant
 def is_speech_available() -> bool:
@@ -2052,6 +2059,12 @@ def is_training_run_on_sagemaker() -> bool:
     return "SAGEMAKER_JOB_NAME" in os.environ
 
 
+# docstyle-ignore
+AUDIO_IMPORT_ERROR = """
+{0} requires soundfile and soxr, which were not found in your environment. You can install them with pip:
+`pip install transformers[audio]`. Please note that you may need to restart your runtime after installation.
+"""
+
 # docstyle-ignore
 AV_IMPORT_ERROR = """
 {0} requires the PyAv library but it was not found in your environment. You can install it with:
@@ -2407,6 +2420,7 @@ def is_training_run_on_sagemaker() -> bool:
 
 BACKENDS_MAPPING = OrderedDict(
     [
+        ("audio", (is_audio_available, AUDIO_IMPORT_ERROR)),
         ("av", (is_av_available, AV_IMPORT_ERROR)),
         ("bs4", (is_bs4_available, BS4_IMPORT_ERROR)),
         ("cv2", (is_cv2_available, CV2_IMPORT_ERROR)),

tests/cli/test_serve.py

@@ -48,10 +48,9 @@
     response_events_to_chunks,
 )
 from transformers.testing_utils import (
+    require_audio,
     require_multipart,
     require_serve,
-    require_soundfile,
-    require_soxr,
     require_torch,
     require_torch_accelerator,
     require_torchcodec,
@@ -2733,8 +2732,7 @@ def serve_kwargs(cls) -> dict:
 
 
 @slow
-@require_soundfile
-@require_soxr
+@require_audio
 @require_multipart
 @require_serve
 class TestTranscription(ServeIntegrationTestCase):

tests/models/muse_spark/test_modeling_muse_spark.py

@@ -29,9 +29,8 @@
 from transformers.models.muse_spark.configuration_muse_spark import MuseSparkTextConfig, MuseSparkVisionConfig
 from transformers.testing_utils import (
     Expectations,
+    require_audio,
     require_deterministic_for_xpu,
-    require_librosa,
-    require_soundfile,
     require_torch,
     require_torch_accelerator,
     require_torch_accelerator_memory,
@@ -536,8 +535,7 @@ def test_integration_video(self):
         )
         self.assertIn(EXPECTED_TEXT_2.get_expectation(), message["content"])
 
-    @require_librosa
-    @require_soundfile
+    @require_audio
     def test_integration_audio(self):
         messages = [
             {

tests/pipelines/test_pipelines_any_to_any.py

@@ -30,8 +30,7 @@
     CaptureLogger,
     Expectations,
     is_pipeline_test,
-    require_soundfile,
-    require_soxr,
+    require_audio,
     require_torch,
     require_vision,
     slow,
@@ -53,8 +52,7 @@
 
 @is_pipeline_test
 @require_vision
-@require_soundfile
-@require_soxr
+@require_audio
 @require_torch
 class AnyToAnyPipelineTests(unittest.TestCase):
     model_mapping = MODEL_FOR_MULTIMODAL_LM_MAPPING

tests/utils/test_audio_utils.py

@@ -45,9 +45,8 @@
 )
 from transformers.testing_utils import (
     is_librosa_available,
+    require_audio,
     require_librosa,
-    require_soundfile,
-    require_soxr,
     require_torchaudio,
     require_torchcodec,
     slow,
@@ -1911,8 +1910,7 @@ def test_format_from_source(self):
 
     # ---- source forms: local path, URL, base64, data: URI -------------------------------------
 
-    @require_soundfile
-    @require_soxr
+    @require_audio
     def test_load_audio_from_every_source_form(self):
         # a single sync gives us the file as a local path, a public URL, raw base64 and a data: URI
         with patch("transformers.audio_utils.is_torchcodec_usable", return_value=False):
@@ -1921,8 +1919,7 @@ def test_load_audio_from_every_source_form(self):
                 self.assertIsInstance(audio, np.ndarray, form)
                 self.assertGreater(audio.size, 0, form)
 
-    @require_soundfile
-    @require_soxr
+    @require_audio
     def test_invalid_base64_raises_value_error(self):
         # a non-URL, non-path string is treated as base64 and must fail cleanly
         with patch("transformers.audio_utils.is_torchcodec_usable", return_value=False):
@@ -1932,8 +1929,7 @@ def test_invalid_base64_raises_value_error(self):
 
     # ---- soundfile fallback (torchcodec unavailable) ------------------------------------------
 
-    @require_soundfile
-    @require_soxr
+    @require_audio
     def test_soundfile_decodes_soundfile_formats(self):
         with patch("transformers.audio_utils.is_torchcodec_usable", return_value=False):
             for name, filetype in self._BUCKET_FILETYPES.items():
@@ -1944,8 +1940,7 @@ def test_soundfile_decodes_soundfile_formats(self):
                 self.assertGreater(audio.size, 0, name)
 
     @require_librosa
-    @require_soundfile
-    @require_soxr
+    @require_audio
     def test_soundfile_matches_librosa_on_soundfile_formats(self):
         # the soundfile backend replaces `librosa.load`, so it must return exactly the same samples
         for name, filetype in self._BUCKET_FILETYPES.items():
@@ -1955,8 +1950,7 @@ def test_soundfile_matches_librosa_on_soundfile_formats(self):
             expected = librosa.load(self._path(name), sr=16000)[0]
             np.testing.assert_array_equal(audio, expected, err_msg=name)
 
-    @require_soundfile
-    @require_soxr
+    @require_audio
     def test_soundfile_rejects_torchcodec_only_formats_with_helpful_error(self):
         with patch("transformers.audio_utils.is_torchcodec_usable", return_value=False):
             for name, filetype in self._BUCKET_FILETYPES.items():
@@ -1980,8 +1974,7 @@ def test_video_url_rejected_up_front_without_download(self):
         self.assertIn("torchcodec", message)
         fetch.assert_not_called()
 
-    @require_soundfile
-    @require_soxr
+    @require_audio
     def test_deprecated_load_audio_librosa_warns_and_pins_soundfile(self):
         # BC: even with torchcodec available, the alias must still decode with soundfile (what librosa used)
         with patch("transformers.audio_utils.is_torchcodec_usable", return_value=True):
@@ -2026,8 +2019,7 @@ def test_unknown_backend_raises(self):
             load_audio(self._path("audio.wav"), sampling_rate=16000, backend="not_a_backend")
 
 
-@require_soundfile
-@require_soxr
+@require_audio
 class LoadAudioSoundfileTester(unittest.TestCase):
     """The `"soundfile"` backend replaces `librosa.load`: same samples, without librosa."""