app

asr.py

@@ -15,7 +15,7 @@ asr_model = asr_model.to(device)
 
 def asr(user_audio):
     # Load and resample user audio
-    audio_input, sampling_rate = util.load_and_resample_audio(file_path = user_audio, target_rate=16000)
+    audio_input, sampling_rate = util.load_and_resample_audio(audio_data = user_audio, target_rate=16000)
     
     # Process audio through ASR model
     inputs = asr_processor(audio_input.squeeze(), sampling_rate=sampling_rate, return_tensors="pt", padding=True)

util.py

@@ -4,6 +4,7 @@ import torchaudio
 import string
 import epitran
 from difflib import SequenceMatcher
+import numpy as np
 
 # Lists of Uyghur short and long texts
 short_texts = [
@@ -12,7 +13,7 @@ short_texts = [
 long_texts = [
     "مەكتەپكە بارغاندا تېخىمۇ بىلىملىك بولۇمەن.",
     "يېزا مەنزىرىسى ھەقىقەتەن گۈزەل.",
-    "بىزنىڭ ئۆيدەپ تۆت تەكچە تۆتىلىسى تەكتەكچە"
+    "بىزنىڭ ئۆيدە تۆت تەكچە تۆتىلىسى تەكتەكچە"
 ]
 
 # Front-End Utils
@@ -35,12 +36,22 @@ def generate_long_text(script_choice):
     return text
 
 # ASR Utils
-def load_and_resample_audio(file_path, target_rate):
+def load_and_resample_audio(audio_data, target_rate):
     """Load audio and resample based on target sample rate"""
-    audio_input, sampling_rate = torchaudio.load(file_path)
+    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)
+
     return audio_input, target_rate
 
 def calculate_pronunciation_accuracy(reference_text, output_text, language_code='uig-Arab'):