Use wider spectrogram when asked

codes/models/gpt_voice/unet_diffusion_tts9.py

@@ -223,8 +223,8 @@ class DiffusionTts(nn.Module):
             ))
         self.latent_converter = nn.Conv1d(in_latent_channels, conditioning_dim, 1)
         self.aligned_latent_padding_embedding = nn.Parameter(torch.randn(1,in_latent_channels,1))
-        if in_channels == 80:
-            self.contextual_embedder = nn.Sequential(nn.Conv1d(80,conditioning_dim,3,padding=1,stride=2),
+        if in_channels > 60:  # It's a spectrogram.
+            self.contextual_embedder = nn.Sequential(nn.Conv1d(in_channels,conditioning_dim,3,padding=1,stride=2),
                                                      CheckpointedXTransformerEncoder(
                                                          needs_permute=True,
                                                          max_seq_len=-1,

codes/scripts/audio/gen_mel.py

@@ -0,0 +1,21 @@
+import os
+
+import torch
+
+from data.util import find_files_of_type, is_audio_file
+from trainer.injectors.audio_injectors import MelSpectrogramInjector
+from utils.util import load_audio
+
+if __name__ == '__main__':
+    path = 'C:\\Users\\jbetk\\Documents\\tmp\\some_audio'
+
+    inj = MelSpectrogramInjector({'in': 'wav', 'out': 'mel',
+                                  'mel_fmax': 12000, 'sampling_rate': 22050, 'n_mel_channels': 100
+                                  },{})
+    audio = find_files_of_type('img', path, qualifier=is_audio_file)[0]
+    for clip in audio:
+        if not clip.endswith('.wav'):
+            continue
+        wav = load_audio(clip, 24000)
+        mel = inj({'wav': wav.unsqueeze(0)})['mel']
+        torch.save(mel, clip.replace('.wav', '.mel'))