diarization1Mæló

app.py

@@ -1,12 +1,12 @@
-# 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
 from transformers import pipeline
 from pyannote.audio import Pipeline
-import torch
 import tempfile
-from torch.serialization import safe_globals  # ← KEY FIX
+import torch
+from torch.serialization import safe_globals
 
 MODEL_NAME = "palli23/whisper-small-sam_spjall"
 
@@ -15,11 +15,8 @@ def transcribe_with_diarization(audio_path):
     if not audio_path:
         return "Hladdu upp hljóðskrá"
     
-    # FIX: Allowlist blocked globals for PyTorch 2.6+
-    with safe_globals([
-        torch.torch_version.TorchVersion,
-        'pyannote.audio.core.task.Specifications'  # Add if needed
-    ]):
+    # LÖGUM PyTorch 2.6+ unpickling villuna
+    with safe_globals([torch.torch_version.TorchVersion]):
         diarization = Pipeline.from_pretrained(
             "pyannote/speaker-diarization-3.1",
             token=os.getenv("HF_TOKEN")
@@ -27,7 +24,6 @@ def transcribe_with_diarization(audio_path):
     
     dia = diarization(audio_path)
     
-    # Whisper-small
     asr = pipeline(
         "automatic-speech-recognition",
         model=MODEL_NAME,
@@ -39,15 +35,23 @@ 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")
-            segment_path = f.name
+            seg = f.name
         
-        text = asr(segment_path)["text"].strip()
+        text = asr(seg)["text"].strip()
         result.append(f"[MÆLENDI {speaker}] {text}")
-        os.unlink(segment_path)
+        os.unlink(seg)
     
     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 · Fixed PyTorch 2
+    gr.Markdown("**Whisper-small + pyannote 3.1 · Keyrir á A100**")
+    gr.Markdown("Fullkominn podcast-transcript með réttum mælendum")
+    
+    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")
+    
+    btn.click(transcribe_with_diarization, audio, out)
+
+demo.launch(auth=("beta", "beta2025"))