diarization

app.py

@@ -1,10 +1,12 @@
-# app.py – Mælendagreining VIRKAR á ZeroGPU (2025 fix)
+# 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 torch
 import tempfile
+from torch.serialization import safe_globals  # ← KEY FIX
 
 MODEL_NAME = "palli23/whisper-small-sam_spjall"
 
@@ -13,13 +15,17 @@ def transcribe_with_diarization(audio_path):
     if not audio_path:
         return "Hladdu upp hljóðskrá"
     
-    # Mælendagreining – 2025 syntax
-    diarization = Pipeline.from_pretrained(
-        "pyannote/speaker-diarization-3.1",
-        token=os.getenv("HF_TOKEN")   # ← FIX
-    ).to("cuda")
+    # 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")
+        ).to("cuda")
     
-    dia_result = diarization(audio_path)
+    dia = diarization(audio_path)
     
     # Whisper-small
     asr = pipeline(
@@ -29,21 +35,22 @@ def transcribe_with_diarization(audio_path):
         token=os.getenv("HF_TOKEN")
     )
     
-    full_text = ""
-    for turn, _, speaker in dia_result.itertracks(yield_label=True):
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
-            dia_result.crop(audio_path, turn).export(tmp.name, format="wav")
-            segment_path = tmp.name
+    result = []
+    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")
+            segment_path = f.name
         
         text = asr(segment_path)["text"].strip()
-        full_text += f"[MÆLENDI {speaker}] {text}\n"
+        result.append(f"[MÆLENDI {speaker}] {text}")
         os.unlink(segment_path)
     
-    return full_text or "Ekkert heyrt"
+    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 · 2025 fix**")
+    gr.Markdown("**Whisper-small + pyannote 3.1 · Fixed PyTorch 2.6+**")
     
     audio = gr.Audio(type="filepath")
     btn = gr.Button("Transcribe með mælendum", variant="primary")