fix tedlium load_audio (#934)

and add test on other backend.

test/torchaudio_unittest/common_utils/backend_utils.py

@@ -6,11 +6,15 @@ import torchaudio
 def set_audio_backend(backend):
     """Allow additional backend value, 'default'"""
     backends = torchaudio.list_audio_backends()
-    if backend == 'default':
+    if backend == 'soundfile-new':
+        be = 'soundfile'
+        torchaudio.USE_SOUNDFILE_LEGACY_INTERFACE = False
+    elif backend == 'default':
         if 'sox_io' in backends:
             be = 'sox_io'
         elif 'soundfile' in backends:
             be = 'soundfile'
+            torchaudio.USE_SOUNDFILE_LEGACY_INTERFACE = True
         else:
             raise unittest.SkipTest('No default backend available')
     else:

test/torchaudio_unittest/datasets/tedlium_test.py

 import os
+import platform
+import unittest
 
 from torchaudio.datasets import tedlium
 
 from torchaudio_unittest.common_utils import (
+    TestBaseMixin,
     TempDirMixin,
     TorchaudioTestCase,
     get_whitenoise,
@@ -30,9 +33,7 @@ PHONEME = [
 ]
 
 
-class TestTedlium(TempDirMixin, TorchaudioTestCase):
-    backend = "default"
-
+class Tedlium(TempDirMixin):
     root_dir = None
     samples = {}
 
@@ -151,3 +152,20 @@ class TestTedlium(TempDirMixin, TorchaudioTestCase):
         phoneme_dict = dataset.phoneme_dict
         phoenemes = [f"{key} {' '.join(value)}" for key, value in phoneme_dict.items()]
         assert phoenemes == PHONEME
+
+
+class TestTedliumSoundfile(Tedlium, TorchaudioTestCase):
+    backend = "soundfile"
+
+
+class TestTedliumSoundfileNew(Tedlium, TorchaudioTestCase):
+    backend = "soundfile-new"
+
+
+if platform.system() != "Windows":
+    class TestTedliumSox(Tedlium, TorchaudioTestCase):
+        backend = "sox"
+
+
+    class TestTedliumSoxIO(Tedlium, TorchaudioTestCase):
+        backend = "sox_io"

torchaudio/datasets/tedlium.py

@@ -143,9 +143,14 @@ class TEDLIUM(Dataset):
         """
         start_time = int(float(start_time) * sample_rate)
         end_time = int(float(end_time) * sample_rate)
-        if torchaudio.get_audio_backend() == "sox_io":
-            return torchaudio.load(path, frame_offset=start_time, num_frames=end_time - start_time)
-        return torchaudio.load(path)[:, start_time:end_time]
+
+        backend = torchaudio.get_audio_backend()
+        if backend == "sox" or (backend == "soundfile" and torchaudio.USE_SOUNDFILE_LEGACY_INTERFACE):
+            kwargs = {"offset": start_time, "num_frames": end_time - start_time}
+        else:
+            kwargs = {"frame_offset": start_time, "num_frames": end_time - start_time}
+
+        return torchaudio.load(path, **kwargs)
 
     def __getitem__(self, n: int) -> Tuple[Tensor, int, str, int, int, int]:
         """Load the n-th sample from the dataset.