fsd

app.py

@@ -20,7 +20,7 @@ import numpy as np
 device = "cuda" if torch.cuda.is_available() else "cpu"
 BATCH_SIZE = 8
 FILE_LIMIT_MB = 1000
-COMPUTE_TYPE = "float16" 
+COMPUTE_TYPE = "float32" 
 YT_LENGTH_LIMIT_S = 3600  # limit to 1 hour YouTube files
 
 num_speakers = 2
@@ -34,9 +34,11 @@ def getpreferredencoding(do_setlocale = True):
 locale.getpreferredencoding = getpreferredencoding
 embedding_model = PretrainedSpeakerEmbedding(
     "speechbrain/spkrec-ecapa-voxceleb",
-    device=torch.device("cuda"))
+    device=torch.device("cpu"))
 model = whisper.load_model(model_size)
 audio = Audio()
+torch.set_default_dtype(torch.float32)
+
 def segment_embedding(segment,duration,path):
     start = segment["start"]
     # Whisper overshoots the end timestamp in the last segment

requirements.txt

@@ -12,3 +12,4 @@ yt-dlp
 more_itertools
 faster-whisper
 git+https://github.com/openai/whisper.git
+gradio_client