LCNN fs 22050

preprocess.py

@@ -15,7 +15,7 @@ def pad_audio(x, max_len=48000):
     return padded_x
 
 
-def preprocess(audio_file):
+def preprocess(audio_file, target_sr=16000, win_dur=4):
     print(f'Preprocessing {audio_file}')
     # Load the audio file
     # y, sr = librosa.load(audio_file, sr=16000)
@@ -26,12 +26,13 @@ def preprocess(audio_file):
     if y.ndim > 1:
         y = np.mean(y, axis=1)
 
-    y = librosa.resample(y, orig_sr=sr_orig, target_sr=16000)
-    sr = 16000
+    if not sr_orig == target_sr:
+        y = librosa.resample(y, orig_sr=sr_orig, target_sr=target_sr)
+    sr = target_sr
 
     # Evaluate N windows of the audio file
     num_eval = 5
-    win_len = int(4*sr)
+    win_len = int(win_dur*sr)
     last_sample = len(y) - win_len
     # start_sample_list = np.linspace(0, max(0, last_sample), num=num_eval)
     start_sample_list = [random.randint(0, max(0, last_sample)) for _ in range(num_eval)]

script.py

@@ -57,7 +57,8 @@ model = LCNN(return_emb=False).to(device)
 # model_path = './checkpoints/LCNN_ALL_DATA.pth'
 # model_path = './checkpoints/LCNN_ALL_DATA_AUG.pth'
 # model_path = './checkpoints/LCNN_ALL_DATA_TTS_AUG.pth'
-model_path = './checkpoints/LCNN_ALL_DATA_TTS_MOD.pth'
+# model_path = './checkpoints/LCNN_ALL_DATA_TTS_MOD.pth'
+model_path = './checkpoints/LCNN_ALL_DATA_HI_FREQ_22050.pth'
 model.load_state_dict(torch.load(model_path, map_location=device))
 
 # # MOE MODEL
@@ -127,7 +128,7 @@ for el in tqdm.tqdm(dataset_remote):
 
     # RUNNING ON HUGGINGFACE
     file_like = io.BytesIO(el["audio"]["bytes"])
-    tensor, sr = preprocess(file_like)
+    tensor, sr = preprocess(file_like, target_sr=22050)
     # # RUNNING LOCALLY
     # tensor = preprocess(el)