Use torchaudio for sox effects

README.md

@@ -17,9 +17,9 @@ Among others, it implements the augmentations that we found to be most useful fo
 Internally, WavAugment uses [libsox](http://sox.sourceforge.net/libsox.html) and allows interleaving of libsox- and pytorch-based effects. 
 
 ### Requirements
- * Linux (WavAugment is not tested under MacOS and might not work properly);
- * [pytorch](pytorch.org) (however, there is also an option of using WavAugment directly from C++ w/o torch, see below);
- * `libsox`: if you have [torchaudio](https://github.com/pytorch/audio) installed, most likely you already have `libsox`. Otherwise, you need to install it, e.g. by running `sudo apt-get install sox libsox-dev libsox-fmt-all`
+ * Linux or MacOS
+ * [pytorch](pytorch.org) >= 1.7
+ * [torchaudio](pytorch.org/audio) >= 0.7
 
 ### Installation
 To install WavAugment, run the following command:
@@ -113,15 +113,6 @@ WavAugment remains explicit and doesn't add effects under the hood.
 If you want to emulate a Sox command that decomposes into several effects, we advise to consult `sox -V` and apply the effects manually. 
 Try it out on some files before running a heavy machine-learning job. 
 
-## But I want to use it from C++
-### Installation
-It is possible to use directly WavAugment's C++ interface to libsox.
-You will need to install `libsox`, e.g. by running
-```bash
-sudo apt-get install sox libsox-dev libsox-fmt-all
-```
-The C++ interface is provided as a single-header library, so you only need to include [this file](./augment/speech_augment.h).
-
 ## Citation
 If you find WavAugment useful in your research, please consider citing:
 ```

augment/effects.py

@@ -7,16 +7,30 @@
 
 import torch
 import numpy as np
-import augment_cpp as _augment
+import torchaudio
+from torchaudio.sox_effects.sox_effects import effect_names as get_effect_names
 
 def shutdown_sox() -> None:
-    _augment.shutdown_sox()
+    pass
 
 
 # Arguments that we can pass to effects
 SoxArg = Optional[List[Union[str, int, Callable]]]
 
 
+class _PyEffectChain:
+
+    def __init__(self):
+        self._effects = []
+
+    def add_effect(self, effect_name, effect_params):
+        params = [str(e) for e in effect_params]
+        self._effects.append([effect_name, *params])
+
+    def apply_flow_effects(self, tensor, src_info, target_info):
+        return torchaudio.sox_effects.apply_effects_tensor(tensor, int(src_info['rate']), self._effects)
+
+
 class SoxEffect:
     def __init__(self, name: str, args: SoxArg):
         self.name = name
@@ -82,18 +96,13 @@ def _apply_sox_effects(chain: List[SoxEffect],
                            src_info: Dict,
                            target_info: Dict) -> Tuple[torch.Tensor, int]:
         instantiated_chain = [x.instantiate() for x in chain]
-        sox_chain = _augment.PyEffectChain()
+        sox_chain = _PyEffectChain()
         for effect_name, effect_args in instantiated_chain:
             sox_chain.add_effect(effect_name, effect_args)
 
-        input_tensor.mul_(EffectChain._NORMALIZER)
-
-        out = input_tensor
-        sr = sox_chain.apply_flow_effects(input_tensor,
-                                          out,
+        out, sr = sox_chain.apply_flow_effects(input_tensor,
                                           src_info,
                                           target_info)
-        out.div_(EffectChain._NORMALIZER)
         return out, sr
 
     def apply(self,
@@ -279,5 +288,5 @@ def create_method(name):
     return lambda s, *x: s._append_effect_to_chain(name, list(x)) # pylint: disable=protected-access
 
 
-for _effect_name in _augment.get_effect_names():
+for _effect_name in get_effect_names():
     setattr(EffectChain, _effect_name, create_method(_effect_name))

requirements.txt

@@ -1,3 +1,3 @@
 torch
-torchaudio
+torchaudio>=0.7
 pytest