added soundfile as primary loader, fixes #845

.travis_dependencies.sh

@@ -33,7 +33,7 @@ if [ ! -d "$src" ]; then
 
         source activate $ENV_NAME
 
-        conda install -c conda-forge ffmpeg
+        conda install -c conda-forge ffmpeg pysoundfile
 
         pip install python-coveralls
 

README.md

@@ -63,15 +63,17 @@ librosa (0.x.x, /path/to/librosa)
 
 ### Hints for the Installation
 
-#### audioread
+#### audioread and MP3 support
 
-*Note that `audioread` needs at least one of the programs to work properly.*
-
-`librosa` uses `audioread` to load audio files.
+`librosa` uses `soundfile` and `audioread` to load audio files.
+Note that `soundfile` does not currently support MP3, which will cause librosa to
+fall back on the `audioread` library.
 
-To fuel `audioread` with more audio-decoding power (e. g. for reading MP3 files),
+To fuel `audioread` with more audio-decoding power (e.g., for reading MP3 files),
 you can either install *ffmpeg* or *GStreamer*.
 
+*Note that `audioread` needs at least one of the programs to work properly.*
+
 If you are using Anaconda, install *ffmpeg* by calling
 ```
 conda install -c conda-forge ffmpeg

librosa/core/audio.py

@@ -5,6 +5,7 @@
 import os
 import six
 
+import soundfile as sf
 import audioread
 import numpy as np
 import scipy.signal
@@ -69,7 +70,7 @@ def load(path, sr=22050, mono=True, offset=0.0, duration=None,
             To use a faster method, set `res_type='kaiser_fast'`.
 
             To use `scipy.signal.resample`, set `res_type='scipy'`.
-            
+
         .. note::
            This uses `audioread`, which may truncate the precision of the
            audio data to 16 bits.
@@ -115,6 +116,40 @@ def load(path, sr=22050, mono=True, offset=0.0, duration=None,
 
     """
 
+    try:
+        with sf.SoundFile(path) as sf_desc:
+            sr_native = sf_desc.samplerate
+            if offset:
+                # Seek to the start of the target read
+                sf_desc.seek(offset * sr_native)
+            if duration:
+                frame_duration = duration * sr_native
+
+            # Load the target number of frames, and transpose to match librosa form
+            y = sf_desc.read(frames=frame_duration, dtype=dtype, always_2d=False).T
+    except:
+        # If soundfile failed, fall back to the audioread loader
+        y, sr_native = __audioread_load(path, offset, duration, dtype)
+
+    # Final cleanup for dtype and contiguity
+    if mono:
+        y = to_mono(y)
+
+    if sr is not None:
+        y = resample(y, sr_native, sr, res_type=res_type)
+
+    else:
+        sr = sr_native
+
+    return y, sr
+
+
+def __audioread_load(path, offset, duration, dtype):
+    '''Load an audio buffer using audioread.
+
+    This loads one block at a time, and then concatenates the results.
+    '''
+
     y = []
     with audioread.audio_open(path) as input_file:
         sr_native = input_file.samplerate
@@ -157,22 +192,12 @@ def load(path, sr=22050, mono=True, offset=0.0, duration=None,
 
     if y:
         y = np.concatenate(y)
-
         if n_channels > 1:
             y = y.reshape((-1, n_channels)).T
-            if mono:
-                y = to_mono(y)
-
-        if sr is not None:
-            y = resample(y, sr_native, sr, res_type=res_type)
-
-        else:
-            sr = sr_native
-
-    # Final cleanup for dtype and contiguity
-    y = np.ascontiguousarray(y, dtype=dtype)
+    else:
+        y = np.empty(0, dtype=dtype)
 
-    return (y, sr)
+    return y, sr_native
 
 
 @cache(level=20)
@@ -380,8 +405,11 @@ def get_duration(y=None, sr=22050, S=None, n_fft=2048, hop_length=512,
     """
 
     if filename is not None:
-        with audioread.audio_open(filename) as fdesc:
-            return fdesc.duration
+        try:
+            return sf.info(filename).duration
+        except:
+            with audioread.audio_open(filename) as fdesc:
+                return fdesc.duration
 
     if y is None:
         if S is None:

setup.py

@@ -53,6 +53,7 @@
         'six >= 1.3',
         'resampy >= 0.2.0',
         'numba >= 0.38.0',
+        'soundfile >= 0.9.0',
     ],
     extras_require={
         'docs': ['numpydoc', 'sphinx!=1.3.1', 'sphinx_rtd_theme',