use_diffuse_tts

codes/models/gpt_voice/unified_voice.py

@@ -151,13 +151,6 @@ class UnifiedGptVoice(nn.Module):
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
 
-    def load_state_dict(self, state_dict: 'OrderedDict[str, Tensor]', strict: bool = True):
-        # Remove the attention biases. I don't know why these are called biases because they are really just fixed attention masks forced into nn.Parameters, which are
-        # easily regenerated and do not need to be saved. This is a hack to allow length modifications and should be removed in the future.
-        filtered = dict(filter(lambda i: not i[0].endswith('.attn.bias'), state_dict.items()))
-        assert len(filtered) == len(state_dict) - len(self.gpt.h)
-        return super().load_state_dict(filtered, strict)
-
     def build_aligned_inputs_and_targets(self, input, start_token, stop_token):
         inp = F.pad(input, (1,0), value=start_token)
         tar = F.pad(input, (0,1), value=stop_token)

codes/scripts/audio/gen/use_diffuse_tts.py

@@ -0,0 +1,59 @@
+import argparse
+import os
+
+import torch
+import torchaudio
+
+from data.audio.unsupervised_audio_dataset import load_audio
+from scripts.audio.gen.speech_synthesis_utils import do_spectrogram_diffusion, \
+    load_discrete_vocoder_diffuser, wav_to_mel, convert_mel_to_codes
+from utils.audio import plot_spectrogram
+from utils.util import load_model_from_config
+
+
+def ceil_multiple(base, multiple):
+    res = base % multiple
+    if res == 0:
+        return base
+    return base + (multiple - res)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model', default='../options/train_diffusion_tts.yml')
+    parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator')
+    parser.add_argument('-diffusion_model_path', type=str, help='Name of the diffusion model in opt.', default='../experiments/train_diffusion_tts\\models\\13600_generator_ema.pth')
+    parser.add_argument('-aligned_codes', type=str, help='Comma-delimited list of integer codes that defines text & prosody. Get this by apply W2V to an existing audio clip or from a bespoke generator.',
+                        default='0,0,0,0,10,10,0,4,0,7,0,17,4,4,0,25,5,0,13,13,0,22,4,4,0,21,15,15,7,0,0,14,4,4,6,8,4,4,0,0,12,5,0,0,5,0,4,4,22,22,8,16,16,0,4,4,4,0,0,0,0,0,0,0')  # Default: 'i am very glad to see you', libritts/train-clean-100/103/1241/103_1241_000017_000001.wav.
+    parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Y:\\clips\\books1\\3042_18_Holden__000000000\\00037.wav')
+    parser.add_argument('-diffusion_steps', type=int, help='Number of diffusion steps to perform to create the generate. Lower steps reduces quality, but >40 is generally pretty good.', default=100)
+    parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='../results/use_diffuse_tts')
+    args = parser.parse_args()
+    os.makedirs(args.output_path, exist_ok=True)
+
+    print("Loading Diffusion Model..")
+    diffusion = load_model_from_config(args.opt, args.diffusion_model_name, also_load_savepoint=False, load_path=args.diffusion_model_path)
+    aligned_codes_compression_factor = 221  # Derived empirically for 11025Hz sample rate. Adjust if sample rate increases.
+
+    print("Loading data..")
+    aligned_codes = torch.tensor([int(s) for s in args.aligned_codes.split(',')]).cuda()
+    diffuser = load_discrete_vocoder_diffuser(desired_diffusion_steps=args.diffusion_steps)
+    cond = load_audio(args.cond, 11025).cuda()
+    if cond.shape[-1] > 44000:
+        cond = cond[:,:44000]
+
+    with torch.no_grad():
+        print("Performing inference..")
+        diffusion.eval()
+
+        # Pad MEL to multiples of 4096//spectrogram_compression_factor
+        msl = aligned_codes.shape[-1]
+        dsl = 4096 // aligned_codes_compression_factor
+        gap = dsl - (msl % dsl)
+        if gap > 0:
+            aligned_codes = torch.nn.functional.pad(aligned_codes, (0, gap))  # This still isn't a perfect multiple, but it's close.
+        output_shape = (1, 1, ceil_multiple(aligned_codes.shape[-1]*aligned_codes_compression_factor, 4096))
+        output = diffuser.p_sample_loop(diffusion, output_shape, model_kwargs={'tokens': aligned_codes.unsqueeze(0),
+                                                                               'conditioning_input': cond.unsqueeze(0)})
+
+        torchaudio.save(os.path.join(args.output_path, 'output.wav'), output.cpu().squeeze(0), 11025)