diarization1Mæló

app.py

@@ -1,27 +1,39 @@
-# app.py – 100 % virkar án pyannote/ECAPA – bara þitt Whisper-small
+# app.py – Whisper-small on ZeroGPU (PRO) – 0.2 RTF
 import os
 import gradio as gr
+import spaces  # Required for @spaces.GPU
 from transformers import pipeline
 
-pipe = pipeline(
-    "automatic-speech-recognition",
-    model="palli23/whisper-small-sam_spjall",
-    device=-1,  # CPU fyrst (ZeroGPU PRO keyrir á GPU)
-    token=os.getenv("HF_TOKEN")
-)
+MODEL_NAME = "palli23/whisper-small-sam_spjall"
 
+# Force ZeroGPU allocation for transcribe (A100/T4)
+@spaces.GPU
 def transcribe(audio):
     if not audio:
         return "Hladdu upp hljóðskrá"
-    return pipe(audio)["text"]
+    
+    # Load pipeline on GPU (cached after first run)
+    pipe = pipeline(
+        "automatic-speech-recognition",
+        model=MODEL_NAME,
+        device=0,  # GPU
+        token=os.getenv("HF_TOKEN")
+    )
+    
+    try:
+        result = pipe(audio)
+        return result["text"]
+    except Exception as e:
+        return f"Villa: {str(e)}"
 
-with gr.Blocks() as demo:
-    gr.Markdown("# Íslenskt ASR – Virkar 100 %")
-    gr.Markdown("**Whisper-small · ~4–5 % WER · Einfalt og stöðugt**")
+# Simple interface
+with gr.Blocks(title="Íslenskt ASR – ZeroGPU GPU") as demo:
+    gr.Markdown("# Íslenskt ASR – ZeroGPU (A100/T4)")
+    gr.Markdown("**Whisper-small · ~4 % WER · 0.2 RTF (15–25 sek fyrir 90 sek hljóð)**")
     
-    audio = gr.Audio(type="filepath")
-    btn = gr.Button("Transcribe")
-    out = gr.Textbox(lines=20)
+    audio = gr.Audio(type="filepath", label="Hladdu upp .mp3 / .wav")
+    btn = gr.Button("Transcribe", variant="primary", size="lg")
+    out = gr.Textbox(lines=25, label="Útskrift")
     
     btn.click(transcribe, audio, out)
 

requirements.txt

@@ -1,6 +1,4 @@
 gradio
 transformers
 torch
-spaces
-librosa
-soundfile
+spaces  # For @spaces.GPU