app

asr.py

@@ -1,5 +1,7 @@
+import numpy as np
 from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
 import torch
+import torchaudio
 from umsc import UgMultiScriptConverter
 import util 
 
@@ -13,9 +15,22 @@ asr_processor.tokenizer.set_target_lang("uig-script_latin")
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 asr_model = asr_model.to(device) 
 
-def asr(user_audio):
+def asr(audio_data, target_rate = 16000):
     # Load and resample user audio
-    audio_input, sampling_rate = util.load_and_resample_audio(audio_data = user_audio, target_rate=16000)
+    if isinstance(audio_data, tuple):
+        # microphone
+        sampling_rate, audio_input = audio_data
+        audio_input = (audio_input / 32768.0).astype(np.float32)
+    elif isinstance(audio_data, str):
+        # file upload
+        audio_input, sampling_rate = torchaudio.load(audio_data)
+    else: 
+        return "<<ERROR: Invalid Audio Input Instance: {}>>".format(type(audio_data))
+    
+    # Resample if needed
+    if sampling_rate != target_rate:
+        resampler = torchaudio.transforms.Resample(sampling_rate, target_rate)
+        audio_input = resampler(audio_input)
     
     # Process audio through ASR model
     inputs = asr_processor(audio_input.squeeze(), sampling_rate=sampling_rate, return_tensors="pt", padding=True)