Update app.py

app.py

@@ -1,4 +1,4 @@
-# app.py — Íslenskt ASR – ZeroGPU + repetition_penalty=1.2 (perfect for your model)
+# app.py — Your original working version + repetition_penalty=1.2 + ngram=3
 
 import os
 os.environ["OMP_NUM_THREADS"] = "1"
@@ -11,10 +11,8 @@ import torch
 import gc
 
 # ——————————————————————————————
-# Model + generation settings (repetition_penalty = 1.2)
+# ZeroGPU worker – model loaded inside
 # ——————————————————————————————
-MODEL_NAME = "palli23/whisper-small-sam_spjall"
-
 @spaces.GPU(duration=180)
 def transcribe_3min(audio_path):
     if not audio_path:
@@ -22,26 +20,24 @@ def transcribe_3min(audio_path):
 
     pipe = pipeline(
         "automatic-speech-recognition",
-        model=MODEL_NAME,
+        model="palli23/whisper-small-sam_spjall",
         torch_dtype=torch.float16,
-        device=0,                     # GPU inside @spaces.GPU
-        token=os.getenv("HF_TOKEN"),  # if you have private model
+        device=0,  # GPU inside @spaces.GPU
     )
 
     result = pipe(
         audio_path,
         chunk_length_s=30,
-        stride_length_s=(6, 0),
         batch_size=8,
-        return_timestamps=False,
+        return_timestamps=False,  # ← no timestamps, as you want
         generate_kwargs={
-            "repetition_penalty": 1.2,     # ← exactly what you want
-            "no_repeat_ngram_size": 3,     # extra safety against loops
+            "repetition_penalty": 1.2,     # ← exactly what you asked for
+            "no_repeat_ngram_size": 3,     # ← exactly what you asked for
             "temperature": 0.0,
         }
     )
 
-    # Clean up memory so ZeroGPU never dies
+    # Clean memory so ZeroGPU lives forever
     del pipe
     gc.collect()
     torch.cuda.empty_cache()
@@ -49,30 +45,25 @@ def transcribe_3min(audio_path):
     return result["text"]
 
 # ——————————————————————————————
-# UI – clean and fast
+# UI – clean and simple
 # ——————————————————————————————
 with gr.Blocks() as demo:
     gr.Markdown("# Íslenskt ASR – 3 mínútur")
     gr.Markdown("**palli23/whisper-small-sam_spjall** · mjög lágur WER · allt að 5 mín hljóð")
     gr.Markdown("**Hafa samband:** pallinr1@protonmail.com")
 
-    audio_in = gr.Audio(
-        type="filepath",
-        label="Hlaðið upp .mp3 / .wav (max ~5 mín)"
-    )
+    audio_in = gr.Audio(type="filepath", label="Hlaðið upp .mp3 / .wav")
     btn = gr.Button("Transcribe", variant="primary", size="lg")
-    output = gr.Textbox(lines=25, label="Útskrift", show_word_timestamps=False)
+    output = gr.Textbox(lines=25, label="Útskrift")
 
     btn.click(fn=transcribe_3min, inputs=audio_in, outputs=output)
 
 # ——————————————————————————————
-# Public Space – no login
+# Public launch
 # ——————————————————————————————
 demo.launch(
     share=True,
     server_name="0.0.0.0",
     server_port=7860,
-    show_error=True,
-    quiet=False,
     auth=None
 )