diarization1Mæló

app.py

@@ -1,12 +1,12 @@
-# app.py – Whisper-small + Mælendagreining (pyannote 3.1) – VIRKAR Á ZeroGPU
+# app.py – FIXED Pyannote UnpicklingError (PyTorch 2.6+ Compatible)
 import os
 import gradio as gr
 import spaces
 from transformers import pipeline
 from pyannote.audio import Pipeline
-import tempfile
 import torch
-from torch.serialization import safe_globals
+import tempfile
+from torch.serialization import safe_globals  # ← KEY FIX
 
 MODEL_NAME = "palli23/whisper-small-sam_spjall"
 
@@ -15,8 +15,11 @@ def transcribe_with_diarization(audio_path):
     if not audio_path:
         return "Hladdu upp hljóðskrá"
     
-    # LÖGUM PyTorch 2.6+ unpickling villuna
-    with safe_globals([torch.torch_version.TorchVersion]):
+    # FIX: Allowlist blocked globals for PyTorch 2.6+
+    with safe_globals([
+        torch.torch_version.TorchVersion,
+        'pyannote.audio.core.task.Specifications'  # Add if needed
+    ]):
         diarization = Pipeline.from_pretrained(
             "pyannote/speaker-diarization-3.1",
             token=os.getenv("HF_TOKEN")
@@ -24,6 +27,7 @@ def transcribe_with_diarization(audio_path):
     
     dia = diarization(audio_path)
     
+    # Whisper-small
     asr = pipeline(
         "automatic-speech-recognition",
         model=MODEL_NAME,
@@ -35,23 +39,21 @@ def transcribe_with_diarization(audio_path):
     for turn, _, speaker in dia.itertracks(yield_label=True):
         with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
             dia.crop(audio_path, turn).export(f.name, format="wav")
-            seg = f.name
+            segment_path = f.name
         
-        text = asr(seg)["text"].strip()
+        text = asr(segment_path)["text"].strip()
         result.append(f"[MÆLENDI {speaker}] {text}")
-        os.unlink(seg)
+        os.unlink(segment_path)
     
     return "\n".join(result) or "Ekkert heyrt"
 
+# Interface
 with gr.Blocks() as demo:
     gr.Markdown("# Íslenskt ASR + Mælendagreining")
-    gr.Markdown("**Whisper-small + pyannote 3.1 · Keyrir á A100**")
-    gr.Markdown("Fullkominn podcast-transcript með réttum mælendum")
+    gr.Markdown("**Whisper-small + pyannote 3.1 · Fixed PyTorch 2.6+**")
     
-    audio = gr.Audio(type="filepath", label="Hladdu upp .mp3 / .wav")
-    btn = gr.Button("Transcribe með mælendum", variant="primary", size="lg")
-    out = gr.Textbox(lines=35, label="Útskrift")
+    audio = gr.Audio(type="filepath")
+    btn = gr.Button("Transcribe með mælendum", variant="primary")
+    out = gr.Textbox(lines=35)
     
-    btn.click(transcribe_with_diarization, audio, out)
-
-demo.launch(auth=("beta", "beta2025"))
+    btn.click(transcribe_with_diarization