fix: ../../TEMPLATE/asr1/pyscripts/audio/format_wav_scp.py

egs2/TEMPLATE/asr1/pyscripts/audio/format_wav_scp.py

@@ -106,9 +106,8 @@ def generator(self):
 
         cached = {}
         for utt, (recodeid, st, et) in self.segments_dict.items():
+            wavpath = self.wav_dict[recodeid]
             if recodeid not in cached:
-                wavpath = self.wav_dict[recodeid]
-
                 if wavpath.endswith("|"):
                     if self.multi_columns:
                         raise RuntimeError(
@@ -117,37 +116,33 @@ def generator(self):
                     # Streaming input e.g. cat a.wav |
                     with kaldiio.open_like_kaldi(wavpath, "rb") as f:
                         with BytesIO(f.read()) as g:
-                            retval = soundfile.read(g)
+                            array, rate = soundfile.read(g)
+
                 else:
                     if self.multi_columns:
-                        retval = soundfile_read(
+                        array, rate = soundfile_read(
                             wavs=wavpath.split(),
                             dtype=None,
                             always_2d=False,
                             concat_axis=1,
                         )
                     else:
-                        retval = soundfile.read(wavpath)
-
-                cached[recodeid] = retval
+                        array, rate = soundfile.read(wavpath)
+                cached[recodeid] = array, rate
 
+            array, rate = cached[recodeid]
             # Keep array until the last query
             recodeid_counter[recodeid] -= 1
             if recodeid_counter[recodeid] == 0:
                 cached.pop(recodeid)
+            # Convert starting time of the segment to corresponding sample number.
+            # If end time is -1 then use the whole file starting from start time.
+            if et != -1:
+                array = array[int(st * rate) : int(et * rate)]
+            else:
+                array = array[int(st * rate) :]
 
-            yield utt, self._return(retval, st, et), None, None
-
-    def _return(self, array, st, et):
-        if isinstance(array, (tuple, list)):
-            array, rate = array
-
-        # Convert starting time of the segment to corresponding sample number.
-        # If end time is -1 then use the whole file starting from start time.
-        if et != -1:
-            return array[int(st * rate) : int(et * rate)], rate
-        else:
-            return array[int(st * rate) :], rate
+            yield utt, (array, rate), None, None
 
 
 def main():
@@ -283,6 +278,7 @@ def generator():
                                 dtype=None,
                                 always_2d=False,
                                 concat_axis=1,
+                                return_subtype=True,
                             )
                         else:
                             with soundfile.SoundFile(wavpath) as sf:

espnet2/fileio/sound_scp.py

@@ -14,6 +14,9 @@ def soundfile_read(
     dtype=None,
     always_2d: bool = False,
     concat_axis: int = 1,
+    start: int = 0,
+    end: int = None,
+    return_subtype: bool = False,
 ) -> Tuple[np.array, int]:
     if isinstance(wavs, str):
         wavs = [wavs]
@@ -24,11 +27,17 @@ def soundfile_read(
     prev_wav = None
     for wav in wavs:
         with soundfile.SoundFile(wav) as f:
-            # for supporting half-precision training
+            f.seek(start)
+            if end is not None:
+                frames = end - start
+            else:
+                frames = -1
             if dtype == "float16":
-                array = f.read(dtype="float32", always_2d=always_2d).astype(dtype)
+                array = f.read(
+                    frames, dtype="float32", always_2d=always_2d,
+                ).astype(dtype)
             else:
-                array = f.read(dtype=dtype, always_2d=always_2d)
+                array = f.read(frames, dtype=dtype, always_2d=always_2d)
             rate = f.samplerate
             subtype = f.subtype
             subtypes.append(subtype)
@@ -61,7 +70,10 @@ def soundfile_read(
     else:
         array = np.concatenate(arrays, axis=concat_axis)
 
-    return array, rate, subtypes
+    if return_subtype:
+        return array, rate, subtypes
+    else:
+        return array, rate
 
 
 class SoundScpReader(collections.abc.Mapping):
@@ -124,7 +136,7 @@ def __init__(
     def __getitem__(self, key) -> Tuple[int, np.ndarray]:
         wavs = self.data[key]
 
-        array, rate, _ = soundfile_read(
+        array, rate = soundfile_read(
             wavs,
             dtype=self.dtype,
             always_2d=self.always_2d,