diarization1Mæló

app.py

@@ -1,4 +1,4 @@
-# app.py – FIXED ECAPA (SpeechBrain Native) + Whisper-small – ZeroGPU
+# app.py – FIXED SpeechBrain Torchaudio Error – ZeroGPU Compatible
 import os
 import gradio as gr
 import spaces
@@ -7,17 +7,21 @@ import numpy as np
 from pydub import AudioSegment
 from sklearn.cluster import AgglomerativeClustering
 from transformers import pipeline
-from speechbrain.inference.speaker import EncoderClassifier  # ← Native SpeechBrain
+from speechbrain.inference.speaker import EncoderClassifier  # Native SpeechBrain
 import torch
 import tempfile
 
+# Skip backend check (fixes Torchaudio 2.9 error)
+import warnings
+warnings.filterwarnings("ignore", message="torchaudio.*")
+
 # ÞITT Whisper-small model
 ASR_MODEL = "palli23/whisper-small-sam_spjall"
 
-# SpeechBrain ECAPA (native – no Transformers error)
+# SpeechBrain ECAPA (native – with backend skip)
 embedder = EncoderClassifier.from_hparams(
     source="speechbrain/spkrec-ecapa-voxceleb",
-    savedir="tmp_ecapa_cache"  # local cache
+    savedir="tmp_ecapa_cache"  # local cache to avoid re-download
 )
 
 # Hlaða ASR á GPU
@@ -51,8 +55,10 @@ def extract_speech_segments(path):
 
 def get_embedding(segment):
     with torch.no_grad():
-        emb = embedder.encode_batch(torch.tensor(segment).unsqueeze(0).float() / 32768.0)
-    return emb.squeeze().numpy()
+        # Normalize to [-1, 1]
+        sig = torch.tensor(segment).float() / 32768.0
+        emb = embedder.encode_batch(sig.unsqueeze(0).to("cuda"))
+    return emb.squeeze().cpu().numpy()
 
 @spaces.GPU(duration=120)
 def transcribe_with_speakers(audio_path):
@@ -85,7 +91,7 @@ def transcribe_with_speakers(audio_path):
 # Interface
 with gr.Blocks() as demo:
     gr.Markdown("# Íslenskt ASR + Mælendagreining (WebRTC + ECAPA)")
-    gr.Markdown("**Whisper-small + SpeechBrain ECAPA · Virkar á ZeroGPU**")
+    gr.Markdown("**Whisper-small + SpeechBrain ECAPA · Fixed Torchaudio 2.9**")
     
     audio = gr.Audio(type="filepath")
     btn = gr.Button("Transcribe með mælendum", variant="primary")

requirements.txt

@@ -1,10 +1,11 @@
 gradio
 transformers
-torch
-spaces
+torch==2.6.0
+torchaudio==2.8.1  # ← FIX: Compatible with SpeechBrain (no list_audio_backends removal)
 speechbrain
 webrtcvad
 pydub
 numpy
 scikit-learn
-soundfile
+soundfile
+spaces