fix(audio): restrict max length for mfcc encoding

gnes/encoder/audio/mfcc.py

@@ -26,10 +26,11 @@
 class MfccEncoder(BaseAudioEncoder):
     batch_size = 64
 
-    def __init__(self, n_mfcc: int = 13, sample_rate: int = 16000, *args, **kwargs):
+    def __init__(self, n_mfcc: int = 13, sample_rate: int = 16000, max_length: int = 100, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.n_mfcc = n_mfcc
         self.sample_rate = sample_rate
+        self.max_length = max_length
 
     @batching
     def encode(self, data: List['np.array'], *args, **kwargs) -> np.ndarray:
@@ -38,10 +39,8 @@ def encode(self, data: List['np.array'], *args, **kwargs) -> np.ndarray:
         mfccs = [np.array(librosa.feature.mfcc(y=audio, sr=self.sample_rate, n_mfcc=self.n_mfcc).T)
                  for audio in data]
 
-        max_lenth = max([len(mf) for mf in mfccs])
-
-        mfccs = [np.concatenate((mf, np.zeros((max_lenth - mf.shape[0], self.n_mfcc), dtype=np.float32)), axis=0)
-                 if mf.shape[0] < max_lenth else mf for mf in mfccs]
+        mfccs = [np.concatenate((mf, np.zeros((self.max_length - mf.shape[0], self.n_mfcc), dtype=np.float32)), axis=0)
+                 if mf.shape[0] < self.max_length else mf[:self.max_length] for mf in mfccs]
         mfccs = [mfcc.reshape((1, -1)) for mfcc in mfccs]
         mfccs = np.squeeze(np.array(mfccs), axis=1)
         return mfccs

gnes/service/indexer.py

@@ -44,6 +44,8 @@ def _handler_index(self, msg: 'gnes_pb2.Message'):
                     offsets += [c.offset_nd for c in d.chunks]
                 elif d.doc_type == gnes_pb2.Document.VIDEO:
                     offsets += [c.offset_1d for c in d.chunks]
+                elif d.doc_type == gnes_pb2.Document.AUDIO:
+                    offsets += [c.offset_1d for c in d.chunks]
                 weights += [c.weight for c in d.chunks]
 
         from ..indexer.base import BaseVectorIndexer, BaseTextIndexer