Update app.py

app.py

@@ -2,6 +2,7 @@ import numpy as np
 import torch
 import gradio as gr
 from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+import librosa
 
 # Load ASR model & processor
 asr_model_id = "Norphel/wav2vec2-large-mms-1b-dzo-colab"
@@ -17,15 +18,33 @@ asr_model.to(device)
 def generate_text(audio):
     if audio is None:
         return "No audio received"
-        
-    sr, data = audio
+    
+    sr, data = audio  # Unpack the tuple
     data = data.astype(np.float32)
-    input_dict = asr_processor(data, sampling_rate=16_000, return_tensors="pt", padding=True)
+    
+    # Resample to 16kHz if the sample rate is not 16kHz
+    target_sr = 16_000
+    if sr != target_sr:
+        data = librosa.resample(data, orig_sr=sr, target_sr=target_sr)
+        sr = target_sr  # Update the sample rate to 16kHz
+    
+    # Check if audio is too short and apply padding
+    min_length = 16000  # 1 second of audio at 16kHz sample rate
+    if len(data) < min_length:
+        padding_length = min_length - len(data)
+        data = np.pad(data, (0, padding_length), mode='constant')
+    
+    # Process the audio with the processor
+    input_dict = asr_processor(data, sampling_rate=sr, return_tensors="pt", padding=True)
+    
+    # Ensure that the input tensor is moved to the correct device (GPU or CPU)
     logits = asr_model(input_dict.input_values.to(device)).logits
     pred_ids = torch.argmax(logits, dim=-1)[0]
 
+    # Decode the prediction
     return asr_processor.decode(pred_ids)
 
+
 input_audio = gr.Audio(
     sources=["microphone"],
     waveform_options=gr.WaveformOptions(