Merge 6dda1a4f32 into fd20265324

gradio_app.py

@@ -74,6 +74,8 @@ class WhisperxModel:
 
     def transcribe(self, audio_path):
         segments = self.model.transcribe(audio_path, batch_size=8)["segments"]
+        for segment in segments:
+            segment['text'] = replace_numbers_with_words(segment['text'])
         return self.align_model.align(segments, audio_path)
 
 
@@ -177,7 +179,7 @@ def align(seed, transcript, audio_path):
     if align_model is None:
         raise gr.Error("Align model not loaded")
     seed_everything(seed)
-
+    transcript = replace_numbers_with_words(transcript).replace("  ", " ").replace("  ", " ")
     fragments = align_segments(transcript, audio_path)
     segments = [{
         "start": float(fragment["begin"]),

models/voicecraft.py

@@ -711,7 +711,7 @@ class VoiceCraft(
         ##################### silence repetition handling #####################
         # prepare the cache placeholder
         # n_layers, 2, bsz, num_heads, src_len, head_dim
-        past = torch.ones([self.args.num_decoder_layers, 2, x.shape[0]], device=x.device, dtype=torch.float32) if kvcache else None
+        past = torch.ones([self.args.num_decoder_layers, 2, x.shape[0]], device=x.device, dtype=torch.float16) if kvcache else None
         # handle multi-span kv-cache
         new_masked_span = False
         
@@ -1011,7 +1011,7 @@ class VoiceCraft(
 
         # prepare the cache placeholder
         # n_layers, 2, bsz, num_heads, src_len, head_dim
-        past = torch.ones([self.args.num_decoder_layers, 2, x.shape[0]], device=x.device, dtype=torch.float32) if kvcache else None
+        past = torch.ones([self.args.num_decoder_layers, 2, x.shape[0]], device=x.device, dtype=torch.float16) if kvcache else None
         # logging.info(f"number of decoder layers: {self.args.num_decoder_layers}")
         # logging.info(f"number of decoder layers: {self.args.num_decoder_layers}")
         # logging.info(f"number of decoder layers: {self.args.num_decoder_layers}")
@@ -1261,7 +1261,7 @@ class VoiceCraft(
 
         # prepare the cache placeholder
         # n_layers, 2, bsz, num_heads, src_len, head_dim
-        past = torch.ones([self.args.num_decoder_layers, 2, x.shape[0]], device=x.device, dtype=torch.float32) if kvcache else None
+        past = torch.ones([self.args.num_decoder_layers, 2, x.shape[0]], device=x.device, dtype=torch.float16) if kvcache else None
         # logging.info(f"number of decoder layers: {self.args.num_decoder_layers}")
         # logging.info(f"number of decoder layers: {self.args.num_decoder_layers}")
         # logging.info(f"number of decoder layers: {self.args.num_decoder_layers}")