diarization1Mæló

app.py

@@ -1,17 +1,28 @@
 # ============================================================
-# app.py – Whisper-small + Pyannote 2.1.1 (ZeroGPU örugg útgáfa)
+# app.py – Whisper-small + Pyannote 3.1 (ZeroGPU örugg)
 # ============================================================
 
 import os
 import gradio as gr
 import spaces
 import tempfile
+import torch
+
 from transformers import pipeline
 from pyannote.audio import Pipeline
+from torch.serialization import add_safe_globals
+
+
+# ================================================
+# Workaround fyrir PyTorch 2.6 weights-only unpickling
+# ================================================
+add_safe_globals({
+    "Specifications": "pyannote.audio.core.task",
+})
 
 
 ASR_MODEL = "palli23/whisper-small-sam_spjall"
-DIAR_MODEL = "pyannote/speaker-diarization"   # <--- ATH: ekki 3.1
+DIAR_MODEL = "pyannote/speaker-diarization-3.1"
 
 
 @spaces.GPU(duration=120)
@@ -22,10 +33,11 @@ def transcribe_with_diarization(audio_path):
 
     # ----------------------------
     # 1. Load diarization pipeline
+    #    (ENGINN token parameter!)
     # ----------------------------
     diarization = Pipeline.from_pretrained(
         DIAR_MODEL,
-        use_auth_token=os.getenv("HF_TOKEN")  # pyannote 2.x notar þetta
+        cache_dir="/home/user/.cache"
     ).to("cuda")
 
     diar = diarization(audio_path)
@@ -63,9 +75,9 @@ def transcribe_with_diarization(audio_path):
 # ------------------------------------------------------------
 with gr.Blocks() as demo:
     gr.Markdown("# 🎙️ Íslenskt ASR + mælendagreining")
-    gr.Markdown("Whisper-small + pyannote 2.1.1 (ZeroGPU örugg útgáfa)")
+    gr.Markdown("Whisper-small + pyannote 3.1 (ZeroGPU örugg útgáfa)")
 
-    audio = gr.Audio(type="filepath", label="Hlaða inn hljóði (.wav or .mp3)")
+    audio = gr.Audio(type="filepath", label="Hlaða inn hljóði (.wav / .mp3)")
     out = gr.Textbox(lines=30, label="Útskrift með mælendum")
 
     btn = gr.Button("Transcribe með mælendum", variant="primary")