diff --git a/TTS/bin/train_glow_tts.py b/TTS/bin/train_glow_tts.py
index 9db2381e..14a20149 100644
--- a/TTS/bin/train_glow_tts.py
+++ b/TTS/bin/train_glow_tts.py
@@ -538,8 +538,16 @@ def main(args):  # pylint: disable=redefined-outer-name
     num_params = count_parameters(model)
     print("\n > Model has {} parameters".format(num_params), flush=True)
 
-    if 'best_loss' not in locals():
+    if args.restore_step == 0 or not args.best_path:
         best_loss = float('inf')
+        print(" > Starting with inf best loss.")
+    else:
+        print(args.best_path)
+        best_loss = torch.load(args.best_path,
+                               map_location='cpu')['model_loss']
+        print(f" > Starting with loaded last best loss {best_loss}.")
+    keep_best = c.get('keep_best', False)
+    keep_after = c.get('keep_after', 10000)  # void if keep_best False
 
     # define dataloaders
     train_loader = setup_loader(ap, 1, is_val=False, verbose=True)
@@ -549,7 +557,8 @@ def main(args):  # pylint: disable=redefined-outer-name
     model = data_depended_init(train_loader, model)
     for epoch in range(0, c.epochs):
         c_logger.print_epoch_start(epoch, c.epochs)
-        train_avg_loss_dict, global_step = train(train_loader, model, criterion, optimizer,
+        train_avg_loss_dict, global_step = train(train_loader, model,
+                                                 criterion, optimizer,
                                                  scheduler, ap, global_step,
                                                  epoch)
         eval_avg_loss_dict = evaluate(eval_loader, model, criterion, ap,
@@ -558,8 +567,9 @@ def main(args):  # pylint: disable=redefined-outer-name
         target_loss = train_avg_loss_dict['avg_loss']
         if c.run_eval:
             target_loss = eval_avg_loss_dict['avg_loss']
-        best_loss = save_best_model(target_loss, best_loss, model, optimizer, global_step, epoch, c.r,
-                                    OUT_PATH)
+        best_loss = save_best_model(target_loss, best_loss, model, optimizer,
+                                    global_step, epoch, c.r, OUT_PATH,
+                                    keep_best=keep_best, keep_after=keep_after)
 
 
 if __name__ == '__main__':
diff --git a/TTS/bin/train_speedy_speech.py b/TTS/bin/train_speedy_speech.py
index a9a83bbf..4e521451 100644
--- a/TTS/bin/train_speedy_speech.py
+++ b/TTS/bin/train_speedy_speech.py
@@ -502,8 +502,16 @@ def main(args):  # pylint: disable=redefined-outer-name
     num_params = count_parameters(model)
     print("\n > Model has {} parameters".format(num_params), flush=True)
 
-    if 'best_loss' not in locals():
+    if args.restore_step == 0 or not args.best_path:
         best_loss = float('inf')
+        print(" > Starting with inf best loss.")
+    else:
+        print(args.best_path)
+        best_loss = torch.load(args.best_path,
+                               map_location='cpu')['model_loss']
+        print(f" > Starting with loaded last best loss {best_loss}.")
+    keep_best = c.get('keep_best', False)
+    keep_after = c.get('keep_after', 10000)  # void if keep_best False
 
     # define dataloaders
     train_loader = setup_loader(ap, 1, is_val=False, verbose=True)
@@ -522,8 +530,8 @@ def main(args):  # pylint: disable=redefined-outer-name
         if c.run_eval:
             target_loss = eval_avg_loss_dict['avg_loss']
         best_loss = save_best_model(target_loss, best_loss, model, optimizer,
-                                    global_step, epoch, c.r,
-                                    OUT_PATH)
+                                    global_step, epoch, c.r, OUT_PATH,
+                                    keep_best=keep_best, keep_after=keep_after)
 
 
 if __name__ == '__main__':
diff --git a/TTS/bin/train_tacotron.py b/TTS/bin/train_tacotron.py
index 0a53f2a1..cdc68c94 100644
--- a/TTS/bin/train_tacotron.py
+++ b/TTS/bin/train_tacotron.py
@@ -581,8 +581,16 @@ def main(args):  # pylint: disable=redefined-outer-name
     num_params = count_parameters(model)
     print("\n > Model has {} parameters".format(num_params), flush=True)
 
-    if 'best_loss' not in locals():
+    if args.restore_step == 0 or not args.best_path:
         best_loss = float('inf')
+        print(" > Starting with inf best loss.")
+    else:
+        print(args.best_path)
+        best_loss = torch.load(args.best_path,
+                               map_location='cpu')['model_loss']
+        print(f" > Starting with loaded last best loss {best_loss}.")
+    keep_best = c.get('keep_best', False)
+    keep_after = c.get('keep_after', 10000)  # void if keep_best False
 
     # define data loaders
     train_loader = setup_loader(ap,
@@ -634,6 +642,8 @@ def main(args):  # pylint: disable=redefined-outer-name
             epoch,
             c.r,
             OUT_PATH,
+            keep_best=keep_best,
+            keep_after=keep_after,
             scaler=scaler.state_dict() if c.mixed_precision else None
         )
 
diff --git a/TTS/bin/train_vocoder_gan.py b/TTS/bin/train_vocoder_gan.py
index 1f2beb70..ecc33288 100644
--- a/TTS/bin/train_vocoder_gan.py
+++ b/TTS/bin/train_vocoder_gan.py
@@ -545,8 +545,16 @@ def main(args):  # pylint: disable=redefined-outer-name
     num_params = count_parameters(model_disc)
     print(" > Discriminator has {} parameters".format(num_params), flush=True)
 
-    if 'best_loss' not in locals():
+    if args.restore_step == 0 or not args.best_path:
         best_loss = float('inf')
+        print(" > Starting with inf best loss.")
+    else:
+        print(args.best_path)
+        best_loss = torch.load(args.best_path,
+                               map_location='cpu')['model_loss']
+        print(f" > Starting with loaded last best loss {best_loss}.")
+    keep_best = c.get('keep_best', False)
+    keep_after = c.get('keep_after', 10000)  # void if keep_best False
 
     global_step = args.restore_step
     for epoch in range(0, c.epochs):
@@ -571,7 +579,10 @@ def main(args):  # pylint: disable=redefined-outer-name
                                     global_step,
                                     epoch,
                                     OUT_PATH,
-                                    model_losses=eval_avg_loss_dict)
+                                    keep_best=keep_best,
+                                    keep_after=keep_after,
+                                    model_losses=eval_avg_loss_dict,
+                                    )
 
 
 if __name__ == '__main__':
diff --git a/TTS/bin/train_vocoder_wavegrad.py b/TTS/bin/train_vocoder_wavegrad.py
index d8dc88e1..7846aae5 100644
--- a/TTS/bin/train_vocoder_wavegrad.py
+++ b/TTS/bin/train_vocoder_wavegrad.py
@@ -393,8 +393,16 @@ def main(args):  # pylint: disable=redefined-outer-name
     num_params = count_parameters(model)
     print(" > WaveGrad has {} parameters".format(num_params), flush=True)
 
-    if 'best_loss' not in locals():
+    if args.restore_step == 0 or not args.best_path:
         best_loss = float('inf')
+        print(" > Starting with inf best loss.")
+    else:
+        print(args.best_path)
+        best_loss = torch.load(args.best_path,
+                               map_location='cpu')['model_loss']
+        print(f" > Starting with loaded last best loss {best_loss}.")
+    keep_best = c.get('keep_best', False)
+    keep_after = c.get('keep_after', 10000)  # void if keep_best False
 
     global_step = args.restore_step
     for epoch in range(0, c.epochs):
@@ -416,6 +424,8 @@ def main(args):  # pylint: disable=redefined-outer-name
             global_step,
             epoch,
             OUT_PATH,
+            keep_best=keep_best,
+            keep_after=keep_after,
             model_losses=eval_avg_loss_dict,
             scaler=scaler.state_dict() if c.mixed_precision else None
         )
diff --git a/TTS/bin/train_vocoder_wavernn.py b/TTS/bin/train_vocoder_wavernn.py
index b4ffe143..44ffef14 100644
--- a/TTS/bin/train_vocoder_wavernn.py
+++ b/TTS/bin/train_vocoder_wavernn.py
@@ -416,8 +416,16 @@ def main(args):  # pylint: disable=redefined-outer-name
     num_parameters = count_parameters(model_wavernn)
     print(" > Model has {} parameters".format(num_parameters), flush=True)
 
-    if "best_loss" not in locals():
-        best_loss = float("inf")
+    if args.restore_step == 0 or not args.best_path:
+        best_loss = float('inf')
+        print(" > Starting with inf best loss.")
+    else:
+        print(args.best_path)
+        best_loss = torch.load(args.best_path,
+                               map_location='cpu')['model_loss']
+        print(f" > Starting with loaded last best loss {best_loss}.")
+    keep_best = c.get('keep_best', False)
+    keep_after = c.get('keep_after', 10000)  # void if keep_best False
 
     global_step = args.restore_step
     for epoch in range(0, c.epochs):
@@ -440,6 +448,8 @@ def main(args):  # pylint: disable=redefined-outer-name
             global_step,
             epoch,
             OUT_PATH,
+            keep_best=keep_best,
+            keep_after=keep_after,
             model_losses=eval_avg_loss_dict,
             scaler=scaler.state_dict() if c.mixed_precision else None
         )
diff --git a/TTS/tts/configs/config.json b/TTS/tts/configs/config.json
index 48f20e8f..5bd249d9 100644
--- a/TTS/tts/configs/config.json
+++ b/TTS/tts/configs/config.json
@@ -1,172 +1,174 @@
-{
-    "model": "Tacotron2",
-    "run_name": "ljspeech-ddc",
-    "run_description": "tacotron2 with DDC and differential spectral loss.",
-
-    // AUDIO PARAMETERS
-    "audio":{
-        // stft parameters
-        "fft_size": 1024,         // number of stft frequency levels. Size of the linear spectogram frame.
-        "win_length": 1024,      // stft window length in ms.
-        "hop_length": 256,       // stft window hop-lengh in ms.
-        "frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
-        "frame_shift_ms": null,  // stft window hop-lengh in ms. If null, 'hop_length' is used.
-
-        // Audio processing parameters
-        "sample_rate": 22050,   // DATASET-RELATED: wav sample-rate.
-        "preemphasis": 0.0,     // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
-        "ref_level_db": 20,     // reference level db, theoretically 20db is the sound of air.
-
-        // Silence trimming
-        "do_trim_silence": true,// enable trimming of slience of audio as you load it. LJspeech (true), TWEB (false), Nancy (true)
-        "trim_db": 60,          // threshold for timming silence. Set this according to your dataset.
-
-        // Griffin-Lim
-        "power": 1.5,           // value to sharpen wav signals after GL algorithm.
-        "griffin_lim_iters": 60,// #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.
-
-        // MelSpectrogram parameters
-        "num_mels": 80,         // size of the mel spec frame.
-        "mel_fmin": 50.0,        // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
-        "mel_fmax": 7600.0,     // maximum freq level for mel-spec. Tune for dataset!!
-        "spec_gain": 1,
-
-        // Normalization parameters
-        "signal_norm": true,    // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
-        "min_level_db": -100,   // lower bound for normalization
-        "symmetric_norm": true, // move normalization to range [-1, 1]
-        "max_norm": 4.0,        // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
-        "clip_norm": true,      // clip normalized values into the range.
-        "stats_path": "/home/erogol/Data/LJSpeech-1.1/scale_stats.npy"    // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by 'compute_statistics.py'. If it is defined, mean-std based notmalization is used and other normalization params are ignored
-    },
-
-    // VOCABULARY PARAMETERS
-    // if custom character set is not defined,
-    // default set in symbols.py is used
-    // "characters":{
-    //     "pad": "_",
-    //     "eos": "~",
-    //     "bos": "^",
-    //     "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
-    //     "punctuations":"!'(),-.:;? ",
-    //     "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ"
-    // },
-
-    // DISTRIBUTED TRAINING
-    "distributed":{
-        "backend": "nccl",
-        "url": "tcp:\/\/localhost:54321"
-    },
-
-    "reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.
-
-    // TRAINING
-    "batch_size": 32,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
-    "eval_batch_size":16,
-    "r": 7,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
-    "gradual_training": [[0, 7, 64], [1, 5, 64], [50000, 3, 32], [130000, 2, 32], [290000, 1, 32]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
-    "mixed_precision": true,     // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.
-
-    // LOSS SETTINGS
-    "loss_masking": true,       // enable / disable loss masking against the sequence padding.
-    "decoder_loss_alpha": 0.5,  // original decoder loss weight. If > 0, it is enabled
-    "postnet_loss_alpha": 0.25, // original postnet loss weight. If > 0, it is enabled
-    "postnet_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
-    "decoder_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
-    "decoder_ssim_alpha": 0.5,     // decoder ssim loss weight. If > 0, it is enabled
-    "postnet_ssim_alpha": 0.25,     // postnet ssim loss weight. If > 0, it is enabled
-    "ga_alpha": 5.0,           // weight for guided attention loss. If > 0, guided attention is enabled.
-    "stopnet_pos_weight": 15.0, // pos class weight for stopnet loss since there are way more negative samples than positive samples.
-
-
-    // VALIDATION
-    "run_eval": true,
-    "test_delay_epochs": 10,  //Until attention is aligned, testing only wastes computation time.
-    "test_sentences_file": null,  // set a file to load sentences to be used for testing. If it is null then we use default english sentences.
-
-    // OPTIMIZER
-    "noam_schedule": false,        // use noam warmup and lr schedule.
-    "grad_clip": 1.0,              // upper limit for gradients for clipping.
-    "epochs": 1000,                // total number of epochs to train.
-    "lr": 0.0001,                  // Initial learning rate. If Noam decay is active, maximum learning rate.
-    "wd": 0.000001,                // Weight decay weight.
-    "warmup_steps": 4000,          // Noam decay steps to increase the learning rate from 0 to "lr"
-    "seq_len_norm": false,         // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.
-
-    // TACOTRON PRENET
-    "memory_size": -1,             // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame.
-    "prenet_type": "original",     // "original" or "bn".
-    "prenet_dropout": false,       // enable/disable dropout at prenet.
-
-    // TACOTRON ATTENTION
-    "attention_type": "original",  // 'original' , 'graves', 'dynamic_convolution'
-    "attention_heads": 4,          // number of attention heads (only for 'graves')
-    "attention_norm": "sigmoid",   // softmax or sigmoid.
-    "windowing": false,            // Enables attention windowing. Used only in eval mode.
-    "use_forward_attn": false,     // if it uses forward attention. In general, it aligns faster.
-    "forward_attn_mask": false,    // Additional masking forcing monotonicity only in eval mode.
-    "transition_agent": false,     // enable/disable transition agent of forward attention.
-    "location_attn": true,         // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
-    "bidirectional_decoder": false,  // use https://arxiv.org/abs/1907.09006. Use it, if attention does not work well with your dataset.
-    "double_decoder_consistency": true,  // use DDC explained here https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency-draft/
-    "ddc_r": 7,                           // reduction rate for coarse decoder.
-
-    // STOPNET
-    "stopnet": true,               // Train stopnet predicting the end of synthesis.
-    "separate_stopnet": true,      // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.
-
-    // TENSORBOARD and LOGGING
-    "print_step": 25,       // Number of steps to log training on console.
-    "tb_plot_step": 100,    // Number of steps to plot TB training figures.
-    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
-    "save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
-    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
-
-    // DATA LOADING
-    "text_cleaner": "phoneme_cleaners",
-    "enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
-    "num_loader_workers": 4,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
-    "num_val_loader_workers": 4,    // number of evaluation data loader processes.
-    "batch_group_size": 4,  //Number of batches to shuffle after bucketing.
-    "min_seq_len": 6,       // DATASET-RELATED: minimum text length to use in training
-    "max_seq_len": 153,     // DATASET-RELATED: maximum text length
-    "compute_input_seq_cache": false,  // if true, text sequences are computed before starting training. If phonemes are enabled, they are also computed at this stage.
-    "use_noise_augment": true,
-
-    // PATHS
-    "output_path": "/home/erogol/Models/LJSpeech/",
-
-    // PHONEMES
-    "phoneme_cache_path": "/home/erogol/Models/phoneme_cache/",  // phoneme computation is slow, therefore, it caches results in the given folder.
-    "use_phonemes": true,           // use phonemes instead of raw characters. It is suggested for better pronounciation.
-    "phoneme_language": "en-us",     // depending on your target language, pick one from  https://github.com/bootphon/phonemizer#languages
-
-    // MULTI-SPEAKER and GST
-    "use_speaker_embedding": false,      // use speaker embedding to enable multi-speaker learning.
-    "use_gst": false,       			    // use global style tokens
-    "use_external_speaker_embedding_file": false, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
-    "external_speaker_embedding_file": "../../speakers-vctk-en.json", // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
-    "gst":	{			                // gst parameter if gst is enabled
-        "gst_style_input": null,        // Condition the style input either on a
-                                        // -> wave file [path to wave] or
-                                        // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}
-                                        // with the dictionary being len(dict) <= len(gst_style_tokens).
-        "gst_embedding_dim": 512,
-        "gst_num_heads": 4,
-        "gst_style_tokens": 10,
-        "gst_use_speaker_embedding": false
-	},
-
-    // DATASETS
-    "datasets":   // List of datasets. They all merged and they get different speaker_ids.
-        [
-            {
-                "name": "ljspeech",
-                "path": "/home/erogol/Data/LJSpeech-1.1/",
-                "meta_file_train": "metadata.csv", // for vtck if list, ignore speakers id in list for train, its useful for test cloning with new speakers
-                "meta_file_val": null
-            }
-        ]
-}
-
+{
+    "model": "Tacotron2",
+    "run_name": "ljspeech-ddc",
+    "run_description": "tacotron2 with DDC and differential spectral loss.",
+
+    // AUDIO PARAMETERS
+    "audio":{
+        // stft parameters
+        "fft_size": 1024,         // number of stft frequency levels. Size of the linear spectogram frame.
+        "win_length": 1024,      // stft window length in ms.
+        "hop_length": 256,       // stft window hop-lengh in ms.
+        "frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
+        "frame_shift_ms": null,  // stft window hop-lengh in ms. If null, 'hop_length' is used.
+
+        // Audio processing parameters
+        "sample_rate": 22050,   // DATASET-RELATED: wav sample-rate.
+        "preemphasis": 0.0,     // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
+        "ref_level_db": 20,     // reference level db, theoretically 20db is the sound of air.
+
+        // Silence trimming
+        "do_trim_silence": true,// enable trimming of slience of audio as you load it. LJspeech (true), TWEB (false), Nancy (true)
+        "trim_db": 60,          // threshold for timming silence. Set this according to your dataset.
+
+        // Griffin-Lim
+        "power": 1.5,           // value to sharpen wav signals after GL algorithm.
+        "griffin_lim_iters": 60,// #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.
+
+        // MelSpectrogram parameters
+        "num_mels": 80,         // size of the mel spec frame.
+        "mel_fmin": 50.0,        // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
+        "mel_fmax": 7600.0,     // maximum freq level for mel-spec. Tune for dataset!!
+        "spec_gain": 1,
+
+        // Normalization parameters
+        "signal_norm": true,    // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
+        "min_level_db": -100,   // lower bound for normalization
+        "symmetric_norm": true, // move normalization to range [-1, 1]
+        "max_norm": 4.0,        // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
+        "clip_norm": true,      // clip normalized values into the range.
+        "stats_path": "/home/erogol/Data/LJSpeech-1.1/scale_stats.npy"    // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by 'compute_statistics.py'. If it is defined, mean-std based notmalization is used and other normalization params are ignored
+    },
+
+    // VOCABULARY PARAMETERS
+    // if custom character set is not defined,
+    // default set in symbols.py is used
+    // "characters":{
+    //     "pad": "_",
+    //     "eos": "~",
+    //     "bos": "^",
+    //     "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
+    //     "punctuations":"!'(),-.:;? ",
+    //     "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ"
+    // },
+
+    // DISTRIBUTED TRAINING
+    "distributed":{
+        "backend": "nccl",
+        "url": "tcp:\/\/localhost:54321"
+    },
+
+    "reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.
+
+    // TRAINING
+    "batch_size": 32,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
+    "eval_batch_size":16,
+    "r": 7,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
+    "gradual_training": [[0, 7, 64], [1, 5, 64], [50000, 3, 32], [130000, 2, 32], [290000, 1, 32]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
+    "mixed_precision": true,     // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.
+
+    // LOSS SETTINGS
+    "loss_masking": true,       // enable / disable loss masking against the sequence padding.
+    "decoder_loss_alpha": 0.5,  // original decoder loss weight. If > 0, it is enabled
+    "postnet_loss_alpha": 0.25, // original postnet loss weight. If > 0, it is enabled
+    "postnet_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
+    "decoder_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
+    "decoder_ssim_alpha": 0.5,     // decoder ssim loss weight. If > 0, it is enabled
+    "postnet_ssim_alpha": 0.25,     // postnet ssim loss weight. If > 0, it is enabled
+    "ga_alpha": 5.0,           // weight for guided attention loss. If > 0, guided attention is enabled.
+    "stopnet_pos_weight": 15.0, // pos class weight for stopnet loss since there are way more negative samples than positive samples.
+
+
+    // VALIDATION
+    "run_eval": true,
+    "test_delay_epochs": 10,  //Until attention is aligned, testing only wastes computation time.
+    "test_sentences_file": null,  // set a file to load sentences to be used for testing. If it is null then we use default english sentences.
+
+    // OPTIMIZER
+    "noam_schedule": false,        // use noam warmup and lr schedule.
+    "grad_clip": 1.0,              // upper limit for gradients for clipping.
+    "epochs": 1000,                // total number of epochs to train.
+    "lr": 0.0001,                  // Initial learning rate. If Noam decay is active, maximum learning rate.
+    "wd": 0.000001,                // Weight decay weight.
+    "warmup_steps": 4000,          // Noam decay steps to increase the learning rate from 0 to "lr"
+    "seq_len_norm": false,         // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.
+
+    // TACOTRON PRENET
+    "memory_size": -1,             // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame.
+    "prenet_type": "original",     // "original" or "bn".
+    "prenet_dropout": false,       // enable/disable dropout at prenet.
+
+    // TACOTRON ATTENTION
+    "attention_type": "original",  // 'original' , 'graves', 'dynamic_convolution'
+    "attention_heads": 4,          // number of attention heads (only for 'graves')
+    "attention_norm": "sigmoid",   // softmax or sigmoid.
+    "windowing": false,            // Enables attention windowing. Used only in eval mode.
+    "use_forward_attn": false,     // if it uses forward attention. In general, it aligns faster.
+    "forward_attn_mask": false,    // Additional masking forcing monotonicity only in eval mode.
+    "transition_agent": false,     // enable/disable transition agent of forward attention.
+    "location_attn": true,         // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
+    "bidirectional_decoder": false,  // use https://arxiv.org/abs/1907.09006. Use it, if attention does not work well with your dataset.
+    "double_decoder_consistency": true,  // use DDC explained here https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency-draft/
+    "ddc_r": 7,                           // reduction rate for coarse decoder.
+
+    // STOPNET
+    "stopnet": true,               // Train stopnet predicting the end of synthesis.
+    "separate_stopnet": true,      // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.
+
+    // TENSORBOARD and LOGGING
+    "print_step": 25,       // Number of steps to log training on console.
+    "tb_plot_step": 100,    // Number of steps to plot TB training figures.
+    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
+    "save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
+    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
+
+    // DATA LOADING
+    "text_cleaner": "phoneme_cleaners",
+    "enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
+    "num_loader_workers": 4,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
+    "num_val_loader_workers": 4,    // number of evaluation data loader processes.
+    "batch_group_size": 4,  //Number of batches to shuffle after bucketing.
+    "min_seq_len": 6,       // DATASET-RELATED: minimum text length to use in training
+    "max_seq_len": 153,     // DATASET-RELATED: maximum text length
+    "compute_input_seq_cache": false,  // if true, text sequences are computed before starting training. If phonemes are enabled, they are also computed at this stage.
+    "use_noise_augment": true,
+
+    // PATHS
+    "output_path": "/home/erogol/Models/LJSpeech/",
+
+    // PHONEMES
+    "phoneme_cache_path": "/home/erogol/Models/phoneme_cache/",  // phoneme computation is slow, therefore, it caches results in the given folder.
+    "use_phonemes": true,           // use phonemes instead of raw characters. It is suggested for better pronounciation.
+    "phoneme_language": "en-us",     // depending on your target language, pick one from  https://github.com/bootphon/phonemizer#languages
+
+    // MULTI-SPEAKER and GST
+    "use_speaker_embedding": false,      // use speaker embedding to enable multi-speaker learning.
+    "use_gst": false,       			    // use global style tokens
+    "use_external_speaker_embedding_file": false, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
+    "external_speaker_embedding_file": "../../speakers-vctk-en.json", // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
+    "gst":	{			                // gst parameter if gst is enabled
+        "gst_style_input": null,        // Condition the style input either on a
+                                        // -> wave file [path to wave] or
+                                        // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}
+                                        // with the dictionary being len(dict) <= len(gst_style_tokens).
+        "gst_embedding_dim": 512,
+        "gst_num_heads": 4,
+        "gst_style_tokens": 10,
+        "gst_use_speaker_embedding": false
+	},
+
+    // DATASETS
+    "datasets":   // List of datasets. They all merged and they get different speaker_ids.
+        [
+            {
+                "name": "ljspeech",
+                "path": "/home/erogol/Data/LJSpeech-1.1/",
+                "meta_file_train": "metadata.csv", // for vtck if list, ignore speakers id in list for train, its useful for test cloning with new speakers
+                "meta_file_val": null
+            }
+        ]
+}
+
diff --git a/TTS/tts/configs/glow_tts_gated_conv.json b/TTS/tts/configs/glow_tts_gated_conv.json
index d34fbaf0..865c6f29 100644
--- a/TTS/tts/configs/glow_tts_gated_conv.json
+++ b/TTS/tts/configs/glow_tts_gated_conv.json
@@ -93,6 +93,8 @@
     "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
     "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
     "apex_amp_level": null,
 
diff --git a/TTS/tts/configs/glow_tts_ljspeech.json b/TTS/tts/configs/glow_tts_ljspeech.json
index 636d9313..6e15de10 100644
--- a/TTS/tts/configs/glow_tts_ljspeech.json
+++ b/TTS/tts/configs/glow_tts_ljspeech.json
@@ -105,6 +105,8 @@
         "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
         "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
         "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+        "keep_best": false,  // If true, keeps all best_models after keep_after steps
+        "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
         "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
         // DATA LOADING
diff --git a/TTS/tts/configs/ljspeech_tacotron2_dynamic_conv_attn.json b/TTS/tts/configs/ljspeech_tacotron2_dynamic_conv_attn.json
index cd5ad8ab..3cf66870 100644
--- a/TTS/tts/configs/ljspeech_tacotron2_dynamic_conv_attn.json
+++ b/TTS/tts/configs/ljspeech_tacotron2_dynamic_conv_attn.json
@@ -1,171 +1,173 @@
-{
-    "model": "Tacotron2",
-    "run_name": "ljspeech-dcattn",
-    "run_description": "tacotron2 with dynamic convolution attention.",
-
-    // AUDIO PARAMETERS
-    "audio":{
-        // stft parameters
-        "fft_size": 1024,         // number of stft frequency levels. Size of the linear spectogram frame.
-        "win_length": 1024,      // stft window length in ms.
-        "hop_length": 256,       // stft window hop-lengh in ms.
-        "frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
-        "frame_shift_ms": null,  // stft window hop-lengh in ms. If null, 'hop_length' is used.
-
-        // Audio processing parameters
-        "sample_rate": 22050,   // DATASET-RELATED: wav sample-rate.
-        "preemphasis": 0.0,     // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
-        "ref_level_db": 20,     // reference level db, theoretically 20db is the sound of air.
-
-        // Silence trimming
-        "do_trim_silence": true,// enable trimming of slience of audio as you load it. LJspeech (true), TWEB (false), Nancy (true)
-        "trim_db": 60,          // threshold for timming silence. Set this according to your dataset.
-
-        // Griffin-Lim
-        "power": 1.5,           // value to sharpen wav signals after GL algorithm.
-        "griffin_lim_iters": 60,// #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.
-
-        // MelSpectrogram parameters
-        "num_mels": 80,         // size of the mel spec frame.
-        "mel_fmin": 50.0,        // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
-        "mel_fmax": 7600.0,     // maximum freq level for mel-spec. Tune for dataset!!
-        "spec_gain": 1,
-
-        // Normalization parameters
-        "signal_norm": true,    // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
-        "min_level_db": -100,   // lower bound for normalization
-        "symmetric_norm": true, // move normalization to range [-1, 1]
-        "max_norm": 4.0,        // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
-        "clip_norm": true,      // clip normalized values into the range.
-        "stats_path": "/home/erogol/Data/LJSpeech-1.1/scale_stats.npy"    // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by 'compute_statistics.py'. If it is defined, mean-std based notmalization is used and other normalization params are ignored
-    },
-
-    // VOCABULARY PARAMETERS
-    // if custom character set is not defined,
-    // default set in symbols.py is used
-    // "characters":{
-    //     "pad": "_",
-    //     "eos": "~",
-    //     "bos": "^",
-    //     "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
-    //     "punctuations":"!'(),-.:;? ",
-    //     "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ"
-    // },
-
-    // DISTRIBUTED TRAINING
-    "distributed":{
-        "backend": "nccl",
-        "url": "tcp:\/\/localhost:54321"
-    },
-
-    "reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.
-
-    // TRAINING
-    "batch_size": 32,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
-    "eval_batch_size":16,
-    "r": 7,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
-    "gradual_training": [[0, 7, 64], [1, 5, 64], [50000, 3, 32], [130000, 2, 32], [290000, 1, 32]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
-    "mixed_precision": true,     // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.
-
-    // LOSS SETTINGS
-    "loss_masking": true,       // enable / disable loss masking against the sequence padding.
-    "decoder_loss_alpha": 0.5,  // original decoder loss weight. If > 0, it is enabled
-    "postnet_loss_alpha": 0.25, // original postnet loss weight. If > 0, it is enabled
-    "postnet_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
-    "decoder_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
-    "decoder_ssim_alpha": 0.5,     // decoder ssim loss weight. If > 0, it is enabled
-    "postnet_ssim_alpha": 0.25,     // postnet ssim loss weight. If > 0, it is enabled
-    "ga_alpha": 0.0,           // weight for guided attention loss. If > 0, guided attention is enabled.
-    "stopnet_pos_weight": 15.0, // pos class weight for stopnet loss since there are way more negative samples than positive samples.
-
-
-    // VALIDATION
-    "run_eval": true,
-    "test_delay_epochs": 10,  //Until attention is aligned, testing only wastes computation time.
-    "test_sentences_file": null,  // set a file to load sentences to be used for testing. If it is null then we use default english sentences.
-
-    // OPTIMIZER
-    "noam_schedule": false,        // use noam warmup and lr schedule.
-    "grad_clip": 1.0,              // upper limit for gradients for clipping.
-    "epochs": 1000,                // total number of epochs to train.
-    "lr": 0.0001,                  // Initial learning rate. If Noam decay is active, maximum learning rate.
-    "wd": 0.000001,                // Weight decay weight.
-    "warmup_steps": 4000,          // Noam decay steps to increase the learning rate from 0 to "lr"
-    "seq_len_norm": false,         // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.
-
-    // TACOTRON PRENET
-    "memory_size": -1,             // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame.
-    "prenet_type": "original",     // "original" or "bn".
-    "prenet_dropout": false,       // enable/disable dropout at prenet.
-
-    // TACOTRON ATTENTION
-    "attention_type": "dynamic_convolution",  // 'original' , 'graves', 'dynamic_convolution'
-    "attention_heads": 4,          // number of attention heads (only for 'graves')
-    "attention_norm": "softmax",   // softmax or sigmoid.
-    "windowing": false,            // Enables attention windowing. Used only in eval mode.
-    "use_forward_attn": false,     // if it uses forward attention. In general, it aligns faster.
-    "forward_attn_mask": false,    // Additional masking forcing monotonicity only in eval mode.
-    "transition_agent": false,     // enable/disable transition agent of forward attention.
-    "location_attn": true,         // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
-    "bidirectional_decoder": false,  // use https://arxiv.org/abs/1907.09006. Use it, if attention does not work well with your dataset.
-    "double_decoder_consistency": false,  // use DDC explained here https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency-draft/
-    "ddc_r": 7,                           // reduction rate for coarse decoder.
-
-    // STOPNET
-    "stopnet": true,               // Train stopnet predicting the end of synthesis.
-    "separate_stopnet": true,      // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.
-
-    // TENSORBOARD and LOGGING
-    "print_step": 25,       // Number of steps to log training on console.
-    "tb_plot_step": 100,    // Number of steps to plot TB training figures.
-    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
-    "save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
-    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
-    "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
-
-    // DATA LOADING
-    "text_cleaner": "phoneme_cleaners",
-    "enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
-    "num_loader_workers": 4,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
-    "num_val_loader_workers": 4,    // number of evaluation data loader processes.
-    "batch_group_size": 4,  //Number of batches to shuffle after bucketing.
-    "min_seq_len": 6,       // DATASET-RELATED: minimum text length to use in training
-    "max_seq_len": 153,     // DATASET-RELATED: maximum text length
-    "compute_input_seq_cache": false,  // if true, text sequences are computed before starting training. If phonemes are enabled, they are also computed at this stage.
-
-    // PATHS
-    "output_path": "/home/erogol/Models/LJSpeech/",
-
-    // PHONEMES
-    "phoneme_cache_path": "/home/erogol/Models/phoneme_cache/",  // phoneme computation is slow, therefore, it caches results in the given folder.
-    "use_phonemes": true,           // use phonemes instead of raw characters. It is suggested for better pronounciation.
-    "phoneme_language": "en-us",     // depending on your target language, pick one from  https://github.com/bootphon/phonemizer#languages
-
-    // MULTI-SPEAKER and GST
-    "use_speaker_embedding": false,      // use speaker embedding to enable multi-speaker learning.
-    "use_gst": false,       			    // use global style tokens
-    "use_external_speaker_embedding_file": false, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
-    "external_speaker_embedding_file": "../../speakers-vctk-en.json", // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
-    "gst":	{			                // gst parameter if gst is enabled
-        "gst_style_input": null,        // Condition the style input either on a
-                                        // -> wave file [path to wave] or
-                                        // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}
-                                        // with the dictionary being len(dict) <= len(gst_style_tokens).
-        "gst_embedding_dim": 512,
-        "gst_num_heads": 4,
-        "gst_style_tokens": 10,
-        "gst_use_speaker_embedding": false
-	},
-
-    // DATASETS
-    "datasets":   // List of datasets. They all merged and they get different speaker_ids.
-        [
-            {
-                "name": "ljspeech",
-                "path": "/home/erogol/Data/LJSpeech-1.1/",
-                "meta_file_train": "metadata.csv", // for vtck if list, ignore speakers id in list for train, its useful for test cloning with new speakers
-                "meta_file_val": null
-            }
-        ]
-}
-
+{
+    "model": "Tacotron2",
+    "run_name": "ljspeech-dcattn",
+    "run_description": "tacotron2 with dynamic convolution attention.",
+
+    // AUDIO PARAMETERS
+    "audio":{
+        // stft parameters
+        "fft_size": 1024,         // number of stft frequency levels. Size of the linear spectogram frame.
+        "win_length": 1024,      // stft window length in ms.
+        "hop_length": 256,       // stft window hop-lengh in ms.
+        "frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
+        "frame_shift_ms": null,  // stft window hop-lengh in ms. If null, 'hop_length' is used.
+
+        // Audio processing parameters
+        "sample_rate": 22050,   // DATASET-RELATED: wav sample-rate.
+        "preemphasis": 0.0,     // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
+        "ref_level_db": 20,     // reference level db, theoretically 20db is the sound of air.
+
+        // Silence trimming
+        "do_trim_silence": true,// enable trimming of slience of audio as you load it. LJspeech (true), TWEB (false), Nancy (true)
+        "trim_db": 60,          // threshold for timming silence. Set this according to your dataset.
+
+        // Griffin-Lim
+        "power": 1.5,           // value to sharpen wav signals after GL algorithm.
+        "griffin_lim_iters": 60,// #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.
+
+        // MelSpectrogram parameters
+        "num_mels": 80,         // size of the mel spec frame.
+        "mel_fmin": 50.0,        // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
+        "mel_fmax": 7600.0,     // maximum freq level for mel-spec. Tune for dataset!!
+        "spec_gain": 1,
+
+        // Normalization parameters
+        "signal_norm": true,    // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
+        "min_level_db": -100,   // lower bound for normalization
+        "symmetric_norm": true, // move normalization to range [-1, 1]
+        "max_norm": 4.0,        // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
+        "clip_norm": true,      // clip normalized values into the range.
+        "stats_path": "/home/erogol/Data/LJSpeech-1.1/scale_stats.npy"    // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by 'compute_statistics.py'. If it is defined, mean-std based notmalization is used and other normalization params are ignored
+    },
+
+    // VOCABULARY PARAMETERS
+    // if custom character set is not defined,
+    // default set in symbols.py is used
+    // "characters":{
+    //     "pad": "_",
+    //     "eos": "~",
+    //     "bos": "^",
+    //     "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
+    //     "punctuations":"!'(),-.:;? ",
+    //     "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ"
+    // },
+
+    // DISTRIBUTED TRAINING
+    "distributed":{
+        "backend": "nccl",
+        "url": "tcp:\/\/localhost:54321"
+    },
+
+    "reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.
+
+    // TRAINING
+    "batch_size": 32,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
+    "eval_batch_size":16,
+    "r": 7,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
+    "gradual_training": [[0, 7, 64], [1, 5, 64], [50000, 3, 32], [130000, 2, 32], [290000, 1, 32]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
+    "mixed_precision": true,     // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.
+
+    // LOSS SETTINGS
+    "loss_masking": true,       // enable / disable loss masking against the sequence padding.
+    "decoder_loss_alpha": 0.5,  // original decoder loss weight. If > 0, it is enabled
+    "postnet_loss_alpha": 0.25, // original postnet loss weight. If > 0, it is enabled
+    "postnet_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
+    "decoder_diff_spec_alpha": 0.25,     // differential spectral loss weight. If > 0, it is enabled
+    "decoder_ssim_alpha": 0.5,     // decoder ssim loss weight. If > 0, it is enabled
+    "postnet_ssim_alpha": 0.25,     // postnet ssim loss weight. If > 0, it is enabled
+    "ga_alpha": 0.0,           // weight for guided attention loss. If > 0, guided attention is enabled.
+    "stopnet_pos_weight": 15.0, // pos class weight for stopnet loss since there are way more negative samples than positive samples.
+
+
+    // VALIDATION
+    "run_eval": true,
+    "test_delay_epochs": 10,  //Until attention is aligned, testing only wastes computation time.
+    "test_sentences_file": null,  // set a file to load sentences to be used for testing. If it is null then we use default english sentences.
+
+    // OPTIMIZER
+    "noam_schedule": false,        // use noam warmup and lr schedule.
+    "grad_clip": 1.0,              // upper limit for gradients for clipping.
+    "epochs": 1000,                // total number of epochs to train.
+    "lr": 0.0001,                  // Initial learning rate. If Noam decay is active, maximum learning rate.
+    "wd": 0.000001,                // Weight decay weight.
+    "warmup_steps": 4000,          // Noam decay steps to increase the learning rate from 0 to "lr"
+    "seq_len_norm": false,         // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.
+
+    // TACOTRON PRENET
+    "memory_size": -1,             // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame.
+    "prenet_type": "original",     // "original" or "bn".
+    "prenet_dropout": false,       // enable/disable dropout at prenet.
+
+    // TACOTRON ATTENTION
+    "attention_type": "dynamic_convolution",  // 'original' , 'graves', 'dynamic_convolution'
+    "attention_heads": 4,          // number of attention heads (only for 'graves')
+    "attention_norm": "softmax",   // softmax or sigmoid.
+    "windowing": false,            // Enables attention windowing. Used only in eval mode.
+    "use_forward_attn": false,     // if it uses forward attention. In general, it aligns faster.
+    "forward_attn_mask": false,    // Additional masking forcing monotonicity only in eval mode.
+    "transition_agent": false,     // enable/disable transition agent of forward attention.
+    "location_attn": true,         // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
+    "bidirectional_decoder": false,  // use https://arxiv.org/abs/1907.09006. Use it, if attention does not work well with your dataset.
+    "double_decoder_consistency": false,  // use DDC explained here https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency-draft/
+    "ddc_r": 7,                           // reduction rate for coarse decoder.
+
+    // STOPNET
+    "stopnet": true,               // Train stopnet predicting the end of synthesis.
+    "separate_stopnet": true,      // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.
+
+    // TENSORBOARD and LOGGING
+    "print_step": 25,       // Number of steps to log training on console.
+    "tb_plot_step": 100,    // Number of steps to plot TB training figures.
+    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
+    "save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
+    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
+    "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
+
+    // DATA LOADING
+    "text_cleaner": "phoneme_cleaners",
+    "enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
+    "num_loader_workers": 4,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
+    "num_val_loader_workers": 4,    // number of evaluation data loader processes.
+    "batch_group_size": 4,  //Number of batches to shuffle after bucketing.
+    "min_seq_len": 6,       // DATASET-RELATED: minimum text length to use in training
+    "max_seq_len": 153,     // DATASET-RELATED: maximum text length
+    "compute_input_seq_cache": false,  // if true, text sequences are computed before starting training. If phonemes are enabled, they are also computed at this stage.
+
+    // PATHS
+    "output_path": "/home/erogol/Models/LJSpeech/",
+
+    // PHONEMES
+    "phoneme_cache_path": "/home/erogol/Models/phoneme_cache/",  // phoneme computation is slow, therefore, it caches results in the given folder.
+    "use_phonemes": true,           // use phonemes instead of raw characters. It is suggested for better pronounciation.
+    "phoneme_language": "en-us",     // depending on your target language, pick one from  https://github.com/bootphon/phonemizer#languages
+
+    // MULTI-SPEAKER and GST
+    "use_speaker_embedding": false,      // use speaker embedding to enable multi-speaker learning.
+    "use_gst": false,       			    // use global style tokens
+    "use_external_speaker_embedding_file": false, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
+    "external_speaker_embedding_file": "../../speakers-vctk-en.json", // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
+    "gst":	{			                // gst parameter if gst is enabled
+        "gst_style_input": null,        // Condition the style input either on a
+                                        // -> wave file [path to wave] or
+                                        // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}
+                                        // with the dictionary being len(dict) <= len(gst_style_tokens).
+        "gst_embedding_dim": 512,
+        "gst_num_heads": 4,
+        "gst_style_tokens": 10,
+        "gst_use_speaker_embedding": false
+	},
+
+    // DATASETS
+    "datasets":   // List of datasets. They all merged and they get different speaker_ids.
+        [
+            {
+                "name": "ljspeech",
+                "path": "/home/erogol/Data/LJSpeech-1.1/",
+                "meta_file_train": "metadata.csv", // for vtck if list, ignore speakers id in list for train, its useful for test cloning with new speakers
+                "meta_file_val": null
+            }
+        ]
+}
+
diff --git a/TTS/tts/configs/speedy_speech_ljspeech.json b/TTS/tts/configs/speedy_speech_ljspeech.json
index bd511470..9f1d3f8b 100644
--- a/TTS/tts/configs/speedy_speech_ljspeech.json
+++ b/TTS/tts/configs/speedy_speech_ljspeech.json
@@ -109,6 +109,8 @@
    "print_eval": false,     // If True, it prints intermediate loss values in evalulation.
    "save_step": 5000,      // Number of training steps expected to save traninpg stats and checkpoints.
    "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+   "keep_best": false,  // If true, keeps all best_models after keep_after steps
+   "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
    "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.:set n
    "mixed_precision": false,
 
diff --git a/TTS/utils/arguments.py b/TTS/utils/arguments.py
index 948c90d3..d05936dc 100644
--- a/TTS/utils/arguments.py
+++ b/TTS/utils/arguments.py
@@ -43,6 +43,11 @@ def parse_arguments(argv):
         type=str,
         help="Model file to be restored. Use to finetune a model.",
         default="")
+    parser.add_argument(
+        "--best_path",
+        type=str,
+        help="Best model file to be used for extracting best loss.",
+        default="")
     parser.add_argument(
         "--config_path",
         type=str,
@@ -67,11 +72,11 @@ def parse_arguments(argv):
     return parser.parse_args()
 
 
-def get_last_checkpoint(path):
-    """Get latest checkpoint from a list of filenames.
+def get_last_models(path):
+    """Get latest checkpoint or/and best model in path.
 
     It is based on globbing for `*.pth.tar` and the RegEx
-    `checkpoint_([0-9]+)`.
+    `(checkpoint|best_model)_([0-9]+)`.
 
     Parameters
     ----------
@@ -81,7 +86,7 @@ def get_last_checkpoint(path):
     Raises
     ------
     ValueError
-        If no checkpoint files are found.
+        If no checkpoint or best_model files are found.
 
     Returns
     -------
@@ -89,22 +94,37 @@ def get_last_checkpoint(path):
         Last checkpoint filename.
 
     """
-    last_checkpoint_num = 0
-    last_checkpoint = None
-    filenames = glob.glob(
-        os.path.join(path, "/*.pth.tar"))
-    for filename in filenames:
-        try:
-            checkpoint_num = int(
-                re.search(r"checkpoint_([0-9]+)", filename).groups()[0])
-            if checkpoint_num > last_checkpoint_num:
-                last_checkpoint_num = checkpoint_num
-                last_checkpoint = filename
-        except AttributeError:  # if there's no match in the filename
-            pass
-    if last_checkpoint is None:
-        raise ValueError(f"No checkpoints in {path}!")
-    return last_checkpoint
+    file_names = glob.glob(os.path.join(path, "*.pth.tar"))
+    last_models = {}
+    last_model_nums = {}
+    for key in ['checkpoint', 'best_model']:
+        last_model_num = 0
+        last_model = None
+        for file_name in file_names:
+            try:
+                model_num = int(re.search(
+                    f"{key}_([0-9]+)", file_name).groups()[0])
+                if model_num > last_model_num:
+                    last_model_num = model_num
+                    last_model = file_name
+            except AttributeError:  # if there's no match in the filename
+                continue
+        last_models[key] = last_model
+        last_model_nums[key] = last_model_num
+
+    # check what models were found
+    if not last_models:
+        raise ValueError(f"No models found in continue path {path}!")
+    elif 'checkpoint' not in last_models:  # no checkpoint just best model
+        last_models['checkpoint'] = last_models['best_model']
+    elif 'best_model' not in last_models:  # no best model
+        # this shouldn't happen, but let's handle it just in case
+        last_models['best_model'] = None
+    # finally check if last best model is more recent than checkpoint
+    elif last_model_nums['best_model'] > last_model_nums['checkpoint']:
+        last_models['checkpoint'] = last_models['best_model']
+
+    return last_models['checkpoint'], last_models['best_model']
 
 
 def process_args(args, model_type):
@@ -143,15 +163,12 @@ def process_args(args, model_type):
         Class that does the TensorBoard loggind.
 
     """
-    if args.continue_path != "":
+    if args.continue_path:
         args.output_path = args.continue_path
         args.config_path = os.path.join(args.continue_path, "config.json")
-        list_of_files = glob.glob(
-            os.path.join(args.continue_path, "*.pth.tar")
-            )  # * means all if need specific format then *.csv
-        args.restore_path = max(list_of_files, key=os.path.getctime)
-        # checkpoint number based continuing
-        # args.restore_path = get_last_checkpoint(args.continue_path)
+        args.restore_path, best_model = get_last_models(args.continue_path)
+        if not args.best_path:
+            args.best_path = best_model
         print(f" > Training continues for {args.restore_path}")
 
     # setup output paths and read configs
@@ -178,7 +195,7 @@ def process_args(args, model_type):
         print("   >  Mixed precision mode is ON")
 
     out_path = args.continue_path
-    if args.continue_path == "":
+    if not out_path:
         out_path = create_experiment_folder(c.output_path, c.run_name,
                                             args.debug)
 
diff --git a/TTS/vocoder/configs/multiband-melgan_and_rwd_config.json b/TTS/vocoder/configs/multiband-melgan_and_rwd_config.json
index 0b751854..b4d42f4b 100644
--- a/TTS/vocoder/configs/multiband-melgan_and_rwd_config.json
+++ b/TTS/vocoder/configs/multiband-melgan_and_rwd_config.json
@@ -138,6 +138,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING
diff --git a/TTS/vocoder/configs/multiband_melgan_config.json b/TTS/vocoder/configs/multiband_melgan_config.json
index 7a5a13e3..af2af8a3 100644
--- a/TTS/vocoder/configs/multiband_melgan_config.json
+++ b/TTS/vocoder/configs/multiband_melgan_config.json
@@ -128,6 +128,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING
diff --git a/TTS/vocoder/configs/multiband_melgan_config_mozilla.json b/TTS/vocoder/configs/multiband_melgan_config_mozilla.json
index 4978d42f..0f133fa7 100644
--- a/TTS/vocoder/configs/multiband_melgan_config_mozilla.json
+++ b/TTS/vocoder/configs/multiband_melgan_config_mozilla.json
@@ -141,6 +141,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING
diff --git a/TTS/vocoder/configs/parallel_wavegan_config.json b/TTS/vocoder/configs/parallel_wavegan_config.json
index fcd765bd..85e659f4 100644
--- a/TTS/vocoder/configs/parallel_wavegan_config.json
+++ b/TTS/vocoder/configs/parallel_wavegan_config.json
@@ -130,6 +130,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING
diff --git a/TTS/vocoder/configs/universal_fullband_melgan.json b/TTS/vocoder/configs/universal_fullband_melgan.json
index fe4433c2..efb6f3cd 100644
--- a/TTS/vocoder/configs/universal_fullband_melgan.json
+++ b/TTS/vocoder/configs/universal_fullband_melgan.json
@@ -124,6 +124,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 25000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING
diff --git a/TTS/vocoder/configs/wavegrad_libritts.json b/TTS/vocoder/configs/wavegrad_libritts.json
index a271ce33..9107d556 100644
--- a/TTS/vocoder/configs/wavegrad_libritts.json
+++ b/TTS/vocoder/configs/wavegrad_libritts.json
@@ -103,6 +103,8 @@
     "print_eval": false,     // If True, it prints loss values for each step in eval run.
     "save_step": 5000,      // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true,     // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": true,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
     // DATA LOADING
diff --git a/TTS/vocoder/configs/wavernn_config.json b/TTS/vocoder/configs/wavernn_config.json
index effb103b..220904c9 100644
--- a/TTS/vocoder/configs/wavernn_config.json
+++ b/TTS/vocoder/configs/wavernn_config.json
@@ -89,6 +89,8 @@
     "print_eval": false, // If True, it prints loss values for each step in eval run.
     "save_step": 25000, // Number of training steps expected to plot training stats on TB and save model checkpoints.
     "checkpoint": true, // If true, it saves checkpoints per "save_step"
+    "keep_best": false,  // If true, keeps all best_models after keep_after steps
+    "keep_after": 10000,    // Global step after which to keep best models if keep_best is true
     "tb_model_param_stats": false, // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
 
 // DATA LOADING
diff --git a/TTS/vocoder/utils/io.py b/TTS/vocoder/utils/io.py
index 5c42dfca..232d972d 100644
--- a/TTS/vocoder/utils/io.py
+++ b/TTS/vocoder/utils/io.py
@@ -1,4 +1,5 @@
 import os
+import glob
 import torch
 import datetime
 import pickle as pickle_tts
@@ -61,12 +62,13 @@ def save_checkpoint(model, optimizer, scheduler, model_disc, optimizer_disc,
                scheduler_disc, current_step, epoch, checkpoint_path, **kwargs)
 
 
-def save_best_model(target_loss, best_loss, model, optimizer, scheduler,
+def save_best_model(current_loss, best_loss, model, optimizer, scheduler,
                     model_disc, optimizer_disc, scheduler_disc, current_step,
-                    epoch, output_folder, **kwargs):
-    if target_loss < best_loss:
-        file_name = 'best_model.pth.tar'
-        checkpoint_path = os.path.join(output_folder, file_name)
+                    epoch, out_path, keep_best=False, keep_after=10000,
+                    **kwargs):
+    if current_loss < best_loss:
+        best_model_name = f'best_model_{current_step}.pth.tar'
+        checkpoint_path = os.path.join(out_path, best_model_name)
         print(" > BEST MODEL : {}".format(checkpoint_path))
         save_model(model,
                    optimizer,
@@ -77,7 +79,21 @@ def save_best_model(target_loss, best_loss, model, optimizer, scheduler,
                    current_step,
                    epoch,
                    checkpoint_path,
-                   model_loss=target_loss,
+                   model_loss=current_loss,
                    **kwargs)
-        best_loss = target_loss
+        # only delete previous if current is saved successfully
+        if not keep_best or (current_step < keep_after):
+            model_names = glob.glob(
+                os.path.join(out_path, 'best_model*.pth.tar'))
+            for model_name in model_names:
+                if os.path.basename(model_name) == best_model_name:
+                    continue
+                os.remove(model_name)
+        # create symlink to best model for convinience
+        link_name = 'best_model.pth.tar'
+        link_path = os.path.join(out_path, link_name)
+        if os.path.islink(link_path) or os.path.isfile(link_path):
+            os.remove(link_path)
+        os.symlink(best_model_name, os.path.join(out_path, link_name))
+        best_loss = current_loss
     return best_loss