diarization1Mæló

app.py

@@ -1,4 +1,4 @@
-# app.py – FIXED Pyannote UnpicklingError (PyTorch 2.6+ Compatible)
+# app.py – Whisper-small + Mælendagreining (pyannote 3.1) – VIRKAR Á ZeroGPU
 import os
 import gradio as gr
 import spaces
@@ -6,7 +6,7 @@ from transformers import pipeline
 from pyannote.audio import Pipeline
 import torch
 import tempfile
-from torch.serialization import safe_globals  # ← KEY FIX
+from torch.serialization import safe_globals
 
 MODEL_NAME = "palli23/whisper-small-sam_spjall"
 
@@ -15,16 +15,17 @@ def transcribe_with_diarization(audio_path):
     if not audio_path:
         return "Hladdu upp hljóðskrá"
     
-    # FIX: Allowlist blocked globals for PyTorch 2.6+
+    # FIX: PyTorch 2.6+ unpickling villu (ZeroGPU krefst þess)
     with safe_globals([
         torch.torch_version.TorchVersion,
-        'pyannote.audio.core.task.Specifications'  # Add if needed
+        'pyannote.audio.core.task.Specifications'
     ]):
         diarization = Pipeline.from_pretrained(
             "pyannote/speaker-diarization-3.1",
             token=os.getenv("HF_TOKEN")
         ).to("cuda")
     
+    # Keyra mælendagreiningu
     dia = diarization(audio_path)
     
     # Whisper-small
@@ -47,13 +48,16 @@ def transcribe_with_diarization(audio_path):
     
     return "\n".join(result) or "Ekkert heyrt"
 
-# Interface
+# Gradio interface
 with gr.Blocks() as demo:
     gr.Markdown("# Íslenskt ASR + Mælendagreining")
-    gr.Markdown("**Whisper-small + pyannote 3.1 · Fixed PyTorch 2.6+**")
+    gr.Markdown("**Whisper-small + pyannote 3.1 · Full podcast-transcript**")
+    gr.Markdown("Hladdu upp .mp3 / .wav (allt að 5 mín)")
     
-    audio = gr.Audio(type="filepath")
-    btn = gr.Button("Transcribe með mælendum", variant="primary")
-    out = gr.Textbox(lines=35)
+    audio = gr.Audio(type="filepath", label="Hljóðskrá")
+    btn = gr.Button("Transcribe með mælendum", variant="primary", size="lg")
+    out = gr.Textbox(lines=35, label="Útskrift með mælendum")
     
-    btn.click(transcribe_with_diarization
+    btn.click(transcribe_with_diarization, inputs=audio, outputs=out)
+
+demo.launch(auth=("beta", "beta2025"))