student

README.md

@@ -35,7 +35,7 @@ put prosody_model.pt To ./bert/prosody_model.pt
 
 put vits_bert_model.pth To ./vits_bert_model.pth
 
-> python vits_infer.py
+> python vits_infer.py --config ./configs/bert_vits.json --model vits_bert_model.pth
 
 ./vits_infer_out have the waves infered, listen !!!
 
@@ -54,7 +54,21 @@ put 000001-010000.txt to ./data/000001-010000.txt
 ![bert_lose](https://user-images.githubusercontent.com/16432329/220883346-c382bea2-1d2f-4a16-b797-2f9e2d2fb639.png)
 
 ### Model compression based on knowledge distillation
-Coming soon~~~
+Student model has 3× speed of teacher model.
+
+To train:
+
+> python train.py -c configs/bert_vits_student.json -m bert_vits_student
+
+To infer, pretrained student model link:https://drive.google.com/file/d/1hTLWYEKH4GV9mQltrMyr3k2UKUo4chdp/view?usp=sharing
+
+Also get studet model at release page.
+
+> python vits_infer.py --config ./configs/bert_vits_student.json --model vits_bert_student.pth
+
+You can use vits_istft as a student model too.
+
+https://github.com/PlayVoice/vits_chinese/tree/vits_istft
 
 ### Another data Link
 https://github.com/PlayVoice/HuaYan_TTS

configs/bert_vits.json

@@ -1,5 +1,7 @@
 {
   "train": {
+    "train_class": "models.SynthesizerTrn",
+    "eval_class": "models.SynthesizerEval",
     "log_interval": 100,
     "eval_interval": 10000,
     "seed": 1234,
@@ -17,8 +19,8 @@
     "c_kl": 1.0
   },
   "data": {
-    "training_files":"filelists/train.txt",
-    "validation_files":"filelists/valid.txt",
+    "training_files": "filelists/train.txt",
+    "validation_files": "filelists/valid.txt",
     "max_wav_value": 32768.0,
     "sampling_rate": 16000,
     "filter_length": 1024,

configs/bert_vits_student.json

@@ -0,0 +1,53 @@
+{
+  "train": {
+    "train_class": "models.SyntStudentTrn",
+    "eval_class": "models.SynthesizerEval",
+    "teacher": "./vits_bert_model.pth",
+    "log_interval": 100,
+    "eval_interval": 10000,
+    "seed": 1234,
+    "epochs": 20000,
+    "learning_rate": 1e-4,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 8,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 12800,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0
+  },
+  "data": {
+    "training_files":"filelists/train.txt",
+    "validation_files":"filelists/valid.txt",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 16000,
+    "filter_length": 1024,
+    "hop_length": 256,
+    "win_length": 1024,
+    "n_mel_channels": 80,
+    "mel_fmin": 0.0,
+    "mel_fmax": null,
+    "add_blank": false,
+    "n_speakers": 0
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 512,
+    "n_heads": 2,
+    "n_layers": 5,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+    "upsample_rates": [8,8,2,2],
+    "upsample_initial_channel": 256,
+    "upsample_kernel_sizes": [16,16,4,4],
+    "n_layers_q": 3,
+    "use_spectral_norm": false
+  }
+}