Update app.py

app.py

@@ -1,16 +1,18 @@
+import os
+
+# Ensure environment variables are set before loading transformers
+os.environ["HF_HOME"] = "/tmp/hf_cache"
+os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache"
+
 from flask import Flask, request, jsonify
 import requests
 import torch
 import librosa
-import os
 from transformers import WhisperProcessor, WhisperForConditionalGeneration
 
-# Explicitly set writable cache directory
-os.environ['HF_HOME'] = '/tmp/hf_cache'
-
 app = Flask(__name__)
 
-# Temporarily using smaller model for faster testing
+# Use a smaller model for faster CPU loading
 model_id = "openai/whisper-base"
 processor = WhisperProcessor.from_pretrained(model_id)
 model = WhisperForConditionalGeneration.from_pretrained(model_id)
@@ -21,45 +23,39 @@ model.to(device)
 forced_decoder_ids = processor.get_decoder_prompt_ids(language="he", task="transcribe")
 
 def transcribe_audio(audio_url):
+    # 1) Download the file to /tmp
     response = requests.get(audio_url)
     audio_path = "/tmp/temp_audio.wav"
     with open(audio_path, "wb") as f:
         f.write(response.content)
 
-    # Load audio
+    # 2) Load audio with librosa
     waveform, sr = librosa.load(audio_path, sr=16000)
 
-    # Safety limit (1 hour)
+    # 3) Truncate to 1 hour max
     max_duration_sec = 3600
     waveform = waveform[:sr * max_duration_sec]
 
-    # Split into smaller chunks
+    # 4) Split into 25-second chunks
     chunk_duration_sec = 25
     chunk_size = sr * chunk_duration_sec
     chunks = [waveform[i : i + chunk_size] for i in range(0, len(waveform), chunk_size)]
 
     partial_text = ""
     for chunk in chunks:
-        inputs = processor(
-            chunk, 
-            sampling_rate=16000, 
-            return_tensors="pt", 
-            padding=True
-        )
+        # Preprocess chunk
+        inputs = processor(chunk, sampling_rate=16000, return_tensors="pt", padding=True)
         input_features = inputs.input_features.to(device)
 
-        # Generate text
+        # Transcribe chunk
         with torch.no_grad():
             predicted_ids = model.generate(
                 input_features, 
                 forced_decoder_ids=forced_decoder_ids
             )
 
-        transcription = processor.batch_decode(
-            predicted_ids, 
-            skip_special_tokens=True
-        )[0]
-
+        # Convert IDs back to text
+        transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
         partial_text += transcription + "\n"
 
     return partial_text.strip()
@@ -71,8 +67,10 @@ def transcribe_endpoint():
     if not audio_url:
         return jsonify({"error": "Missing 'audio_url' in request"}), 400
 
+    # Perform transcription
     transcription = transcribe_audio(audio_url)
     return jsonify({"transcription": transcription})
 
 if __name__ == '__main__':
+    # Run the Flask app on port 7860
     app.run(host="0.0.0.0", port=7860)