try new approach

app.py

@@ -1,21 +1,25 @@
 import streamlit as st
-from transformers import pipeline
+from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
 import tempfile
 from pydub import AudioSegment
 import numpy as np
+import torch
 
-# Load the ASR pipeline
+
+# Load the processor and model
 @st.cache_resource
-def load_asr_pipeline():
-    asr_pipeline = pipeline("automatic-speech-recognition", model="Yehor/whisper-small-ukrainian")
-    return asr_pipeline
+def load_model_and_processor():
+    processor = AutoProcessor.from_pretrained("arampacha/whisper-large-uk-2")
+    model = AutoModelForSpeechSeq2Seq.from_pretrained("arampacha/whisper-large-uk-2")
+    return processor, model
+
 
 st.title("Voice Recognition App using Whisper")
 
 st.write("Upload an audio file and the Whisper model will transcribe it to text.")
 
-# Load the ASR pipeline
-asr_pipeline = load_asr_pipeline()
+# Load the processor and model
+processor, model = load_model_and_processor()
 st.write("Model loaded successfully.")
 
 # File uploader for audio file
@@ -40,9 +44,18 @@ if uploaded_file is not None:
     audio_input = AudioSegment.from_file(temp_wav_path).set_frame_rate(16000).set_channels(1)
     audio_input = np.array(audio_input.get_array_of_samples(), dtype=np.float32)
 
-    # Perform transcription
-    result = asr_pipeline(audio_input)
+    # Normalize audio
+    audio_input = (audio_input - np.mean(audio_input)) / np.std(audio_input)
+    audio_input = torch.tensor(audio_input).unsqueeze(0)
+
+    # Process the audio
+    input_features = processor(audio_input, sampling_rate=16000, return_tensors="pt").input_features
+
+    # Generate transcription
+    with torch.no_grad():
+        predicted_ids = model.generate(input_features)
+
+    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
 
-    # Display transcription
     st.write("Transcription:")
-    st.write(result['text'])
+    st.write(transcription)