Update app.py

app.py

@@ -15,33 +15,32 @@ from transformers import WhisperProcessor, WhisperForConditionalGeneration
 
 app = Flask(__name__)
 
-# Choose a multilingual Whisper model that includes Hebrew.
-# For CPU usage, 'openai/whisper-base' or 'openai/whisper-tiny' are typical.
-model_id = "openai/whisper-base"
+# Use your custom Hebrew Whisper model
+model_id = "ivrit-ai/whisper-large-v3-turbo"
 processor = WhisperProcessor.from_pretrained(model_id)
 model = WhisperForConditionalGeneration.from_pretrained(model_id)
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 model.to(device)
 
-# Force Hebrew to avoid short-audio meltdown with auto-detect
+# Force Hebrew so it won't require 30s audio for language detection
 forced_decoder_ids = processor.get_decoder_prompt_ids(language="he", task="transcribe")
 
 def transcribe_audio(audio_url):
-    # 1) Download audio to /tmp
+    # 1) Download audio 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)
 
-    # 2) Load audio with librosa
+    # 2) Load with librosa
     waveform, sr = librosa.load(audio_path, sr=16000)
 
-    # 3) Limit up to 1 hour for stability
+    # 3) (Optional) limit up to 1 hour
     max_sec = 3600
     waveform = waveform[: sr * max_sec]
 
-    # 4) Chunk the audio in 25-second intervals
+    # 4) Split into 25-second chunks
     chunk_sec = 25
     chunk_size = sr * chunk_sec
     chunks = [waveform[i : i + chunk_size] for i in range(0, len(waveform), chunk_size)]
@@ -52,7 +51,7 @@ def transcribe_audio(audio_url):
         inputs = processor(chunk, sampling_rate=sr, return_tensors="pt", padding=True)
         input_features = inputs.input_features.to(device)
 
-        # Force Hebrew to skip auto-detect logic
+        # Force Hebrew, skipping auto-detect
         with torch.no_grad():
             predicted_ids = model.generate(
                 input_features,
@@ -71,10 +70,10 @@ def transcribe_endpoint():
     if not audio_url:
         return jsonify({"error": "Missing 'audio_url' in request"}), 400
 
-    # Perform forced-Hebrew transcription
+    # Perform forced Hebrew transcription
     text = transcribe_audio(audio_url)
 
-    # Return JSON with no ASCII escaping (ensures real Hebrew chars)
+    # Return raw Hebrew in JSON
     payload = {"Transcription": text}
     return Response(
         json.dumps(payload, ensure_ascii=False),