replace soundfile with librosa

CHANGELOG.md

@@ -58,6 +58,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Added support for the `ObjectDetector` with FiftyOne ([#727](https://github.com/PyTorchLightning/lightning-flash/pull/727))
 
+- Added support for MP3 files to the `SpeechRecognition` task with librosa ([#726](https://github.com/PyTorchLightning/lightning-flash/pull/726))
+
 ### Changed
 
 - Changed how pretrained flag works for loading weights for ImageClassifier task ([#560](https://github.com/PyTorchLightning/lightning-flash/pull/560))

flash/audio/speech_recognition/data.py

@@ -35,7 +35,7 @@
 from flash.core.utilities.imports import _AUDIO_AVAILABLE, requires_extras
 
 if _AUDIO_AVAILABLE:
-    import soundfile as sf
+    import librosa
     from datasets import Dataset as HFDataset
     from datasets import load_dataset
     from transformers import Wav2Vec2CTCTokenizer
@@ -48,7 +48,7 @@ def deserialize(self, sample: Any) -> Dict:
         encoded_with_padding = (sample + "===").encode("ascii")
         audio = base64.b64decode(encoded_with_padding)
         buffer = io.BytesIO(audio)
-        data, sampling_rate = sf.read(buffer)
+        data, sampling_rate = librosa.load(buffer, sr=16000)
         return {
             DefaultDataKeys.INPUT: data,
             DefaultDataKeys.METADATA: {"sampling_rate": sampling_rate},
@@ -69,7 +69,7 @@ def _load_sample(self, sample: Dict[str, Any]) -> Any:
             and "root" in sample[DefaultDataKeys.METADATA]
         ):
             path = os.path.join(sample[DefaultDataKeys.METADATA]["root"], path)
-        speech_array, sampling_rate = sf.read(path)
+        speech_array, sampling_rate = librosa.load(path, sr=16000)
         sample[DefaultDataKeys.INPUT] = speech_array
         sample[DefaultDataKeys.METADATA] = {"sampling_rate": sampling_rate}
         return sample
@@ -129,7 +129,7 @@ def load_data(self, data: Dataset, dataset: Optional[Any] = None) -> Union[Seque
 
 class SpeechRecognitionPathsDataSource(PathsDataSource, BaseSpeechRecognition):
     def __init__(self):
-        super().__init__(("wav", "ogg", "flac", "mat"))
+        super().__init__(("wav", "ogg", "flac", "mat", "mp3"))
 
     def load_sample(self, sample: Dict[str, Any], dataset: Any = None) -> Any:
         return self._load_sample(sample)

flash/core/utilities/imports.py

@@ -88,7 +88,7 @@ def _compare_version(package: str, op, version) -> bool:
 _PIL_AVAILABLE = _module_available("PIL")
 _OPEN3D_AVAILABLE = _module_available("open3d")
 _SEGMENTATION_MODELS_AVAILABLE = _module_available("segmentation_models_pytorch")
-_SOUNDFILE_AVAILABLE = _module_available("soundfile")
+_LIBROSA_AVAILABLE = _module_available("librosa")
 _TORCH_SCATTER_AVAILABLE = _module_available("torch_scatter")
 _TORCH_SPARSE_AVAILABLE = _module_available("torch_sparse")
 _TORCH_GEOMETRIC_AVAILABLE = _module_available("torch_geometric")
@@ -148,7 +148,7 @@ class Image(metaclass=MetaImage):
 )
 _SERVE_AVAILABLE = _FASTAPI_AVAILABLE and _PYDANTIC_AVAILABLE and _CYTOOLZ_AVAILABLE and _UVICORN_AVAILABLE
 _POINTCLOUD_AVAILABLE = _OPEN3D_AVAILABLE and _TORCHVISION_AVAILABLE
-_AUDIO_AVAILABLE = all([_TORCHAUDIO_AVAILABLE, _SOUNDFILE_AVAILABLE, _TRANSFORMERS_AVAILABLE])
+_AUDIO_AVAILABLE = all([_TORCHAUDIO_AVAILABLE, _LIBROSA_AVAILABLE, _TRANSFORMERS_AVAILABLE])
 _GRAPH_AVAILABLE = _TORCH_SCATTER_AVAILABLE and _TORCH_SPARSE_AVAILABLE and _TORCH_GEOMETRIC_AVAILABLE
 
 _EXTRAS_AVAILABLE = {

requirements/datatype_audio.txt

@@ -1,4 +1,4 @@
 torchaudio
-soundfile>=0.10.2
+librosa>=0.8.1
 transformers>=4.5
 datasets>=1.8