diff --git a/codes/data/audio/paired_voice_audio_dataset.py b/codes/data/audio/paired_voice_audio_dataset.py
index 54cda975..8be74604 100644
--- a/codes/data/audio/paired_voice_audio_dataset.py
+++ b/codes/data/audio/paired_voice_audio_dataset.py
@@ -6,6 +6,7 @@ import torch
 import torch.nn.functional as F
 import torch.utils.data
 import torchaudio
+from tokenizers import Tokenizer
 from tqdm import tqdm
 from transformers import GPT2TokenizerFast
 
@@ -84,7 +85,7 @@ class TextWavLoader(torch.utils.data.Dataset):
         self.needs_collate = opt_get(hparams, ['needs_collate'], True)
         if not self.needs_collate:
             assert self.max_wav_len is not None and self.max_text_len is not None
-        self.tokenizer = GPT2TokenizerFast.from_pretrained('gpt2')
+        self.tokenizer = Tokenizer.from_file(opt_get(hparams, ['tokenizer_vocab'], '../experiments/gpt_tts_tokenizer.json'))
 
     def get_wav_text_pair(self, audiopath_and_text):
         # separate filename and text
@@ -94,7 +95,11 @@ class TextWavLoader(torch.utils.data.Dataset):
         return (text_seq, wav, text, audiopath_and_text[0])
 
     def get_text(self, text):
-        return torch.IntTensor(self.tokenizer(text)['input_ids'])
+        tokens = self.tokenizer.encode(text).ids
+        tokens = torch.IntTensor(tokens)
+        assert not torch.any(tokens == 0)
+        assert not torch.any(tokens == 9999)
+        return tokens
 
     def load_conditioning_candidates(self, path):
         candidates = find_files_of_type('img', os.path.dirname(path), qualifier=is_audio_file)[0]
diff --git a/codes/data/audio/voice_tokenizer_builder.py b/codes/data/audio/voice_tokenizer_builder.py
new file mode 100644
index 00000000..df8aedb7
--- /dev/null
+++ b/codes/data/audio/voice_tokenizer_builder.py
@@ -0,0 +1,47 @@
+from tokenizers import Tokenizer
+from tokenizers.models import BPE
+from tokenizers.pre_tokenizers import Whitespace
+from tokenizers.trainers import BpeTrainer
+
+from data.audio.paired_voice_audio_dataset import load_mozilla_cv, load_voxpopuli, load_tsv
+from models.tacotron2.taco_utils import load_filepaths_and_text
+
+
+def build_text_file_from_priors(priors, output):
+    with open(output, 'w', encoding='utf-8') as out:
+        for p, fm in priors:
+            if fm == 'lj' or fm == 'libritts':
+                fetcher_fn = load_filepaths_and_text
+            elif fm == 'tsv':
+                fetcher_fn = load_tsv
+            elif fm == 'mozilla_cv':
+                fetcher_fn = load_mozilla_cv
+            elif fm == 'voxpopuli':
+                fetcher_fn = load_voxpopuli
+            else:
+                raise NotImplementedError()
+            apt = fetcher_fn(p)
+            for path, text in apt:
+                out.write(text + "\n")
+            out.flush()
+
+
+def train():
+    trainer = BpeTrainer(special_tokens=['[STOP]', '[UNK]'], vocab_size=9999)
+    tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
+    tokenizer.pre_tokenizer = Whitespace()
+    tokenizer.train(['all_texts.txt'], trainer)
+    tokenizer.save('gpt_tts_tokenizer.json')
+
+
+if __name__ == '__main__':
+    '''
+    build_text_file_from_priors([('Y:\\bigasr_dataset\\libritts\\train-all.txt', 'libritts'),
+                                 ('Y:\\bigasr_dataset\\libritts\\test-clean_list.txt', 'libritts'),
+                                 #('Y:\\bigasr_dataset\\voxpopuli\\audio\\transcribed_data\\en\\asr_en.tsv', 'voxpopuli'),
+                                 ('Y:\\bigasr_dataset\\voxpopuli\\audio\\transcribed_data\\en\\asr_train.tsv', 'voxpopuli'),
+                                 ('Y:\\clips\\books1-transcribed.tsv', 'tsv'),
+                                 ('Y:\\clips\\books2-transcribed.tsv', 'tsv'),
+                                 ('Y:\\clips\\podcasts-0-transcribed.tsv', 'tsv')], 'all_texts.txt')
+    '''
+    train()
\ No newline at end of file
diff --git a/codes/models/gpt_voice/gpt_tts_hf.py b/codes/models/gpt_voice/gpt_tts_hf.py
index b804d0d8..54bf6ffe 100644
--- a/codes/models/gpt_voice/gpt_tts_hf.py
+++ b/codes/models/gpt_voice/gpt_tts_hf.py
@@ -39,16 +39,18 @@ class ConditioningEncoder(nn.Module):
 
 
 class GptTtsHf(nn.Module):
-    NUMBER_TEXT_TOKENS = 50257  # The number of BPE tokens produced by the HF GPT2Tokenizer
-    START_TEXT_TOKEN = 50256
+    NUMBER_TEXT_TOKENS = 10000  # The number of tokens produced by our bespoke BPE tokenizer.
+    START_TEXT_TOKEN = 9999
     STOP_TEXT_TOKEN = 0
     NUMBER_MEL_CODES = 8194
     START_MEL_TOKEN = 8192
     STOP_MEL_TOKEN = 8193
 
-    def __init__(self, layers=8, model_dim=512, heads=8, max_symbols_per_phrase=100, max_mel_tokens=250, max_conditioning_inputs=3,
+    def __init__(self, layers=8, model_dim=512, heads=8, max_symbols_per_phrase=80, max_mel_tokens=250, max_conditioning_inputs=3,
                  checkpointing=True, mel_length_compression=1024, max_conditioning_length=60):
         super().__init__()
+
+
         self.max_mel_tokens = max_mel_tokens
         self.max_symbols_per_phrase = max_symbols_per_phrase
         self.model_dim = model_dim