f

codes/scripts/audio/preparation/combine_phonetic_and_text.py

@@ -0,0 +1,32 @@
+import os
+
+if __name__ == '__main__':
+    basepath = 'Y:/clips/books2'
+
+    english_file = os.path.join(basepath, 'transcribed-oco-realtext.tsv')
+    if not os.path.exists(english_file):
+        english_file = os.path.join(basepath, 'transcribed-oco.tsv')
+    phoneme_file = os.path.join(basepath, 'transcribed-phoneme-oco.tsv')
+
+    texts = {}
+    with open(english_file, 'r', encoding='utf-8') as f:
+        for line in f.readlines():
+            spl = line.split('\t')
+            if len(spl) == 3:
+                text, p, _ = spl
+                texts[p] = text
+            else:
+                print(f'Error processing line {line}')
+
+    with open(phoneme_file, 'r', encoding='utf-8') as f:
+        wf = open(os.path.join(basepath, 'transcribed-phoneme-english-oco.tsv'), 'w', encoding='utf-8')
+        for line in f.readlines():
+            spl = line.split('\t')
+            if len(spl) == 3:
+                _, p, codes = spl
+                codes = codes.strip()
+                if p not in texts:
+                    print(f'Could not find the text for {p}')
+                    continue
+                wf.write(f'{texts[p]}\t{p}\t{codes}\n')
+        wf.close()

codes/trainer/eval/audio_diffusion_fid.py

@@ -1,5 +1,7 @@
 import os
 import os.path as osp
+import random
+
 import torch
 import torchaudio
 import torchvision.utils
@@ -33,6 +35,13 @@ class AudioDiffusionFid(evaluator.Evaluator):
         super().__init__(model, opt_eval, env, uses_all_ddp=True)
         self.real_path = opt_eval['eval_tsv']
         self.data = load_tsv_aligned_codes(self.real_path)
+
+        # Deterministically shuffle the data.
+        ostate = random.getstate()
+        random.seed(5)
+        random.shuffle(self.data)
+        random.setstate(ostate)
+
         if 'clip_dataset' in opt_eval.keys():
             self.data = self.data[:opt_eval['clip_dataset']]
         if distributed.is_initialized() and distributed.get_world_size() > 1:

codes/trainer/eval/eval_wer.py

@@ -1,6 +1,6 @@
 from copy import deepcopy
 
-from datasets import load_metric
+#from datasets import load_metric
 
 import torch
 from tqdm import tqdm