use reduction factor of 1.

.gitattributes

@@ -25,7 +25,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
-pretrained_model_ljs_500k.ckpt filter=lfs diff=lfs merge=lfs -text
 bazelisk-linux-amd64 filter=lfs diff=lfs merge=lfs -text
 wavegru.ckpt filter=lfs diff=lfs merge=lfs -text
 wavegru_mod.so filter=lfs diff=lfs merge=lfs -text
+pretrained_model_ljs_570k.ckpt filter=lfs diff=lfs merge=lfs -text

app.py

@@ -10,7 +10,7 @@ from inference import load_tacotron_model, load_wavegru_net, mel_to_wav, text_to
 from wavegru_cpp import extract_weight_mask, load_wavegru_cpp
 
 alphabet, tacotron_net, tacotron_config = load_tacotron_model(
-    "./alphabet.txt", "./tacotron.toml", "./pretrained_model_ljs_500k.ckpt"
+    "./alphabet.txt", "./tacotron.toml", "./pretrained_model_ljs_570k.ckpt"
 )
 
 

inference.py

@@ -28,7 +28,7 @@ def load_tacotron_model(alphabet_file, config_file, model_file):
     return alphabet, net, config
 
 
-tacotron_inference_fn = pax.pure(lambda net, text: net.inference(text, max_len=1200))
+tacotron_inference_fn = pax.pure(lambda net, text: net.inference(text, max_len=2400))
 
 
 def text_to_mel(net, text, alphabet, config):

pretrained_model_ljs_500k.ckpt → pretrained_model_ljs_570k.ckpt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4eabdcac35cd016469d17678f9549bd25d1c9bf66c9089ea9f0632619ba91194
-size 53221435
+oid sha256:ed5e35d9f290c45e229fe8135218e9e626eb3fdb4e48909721752e228f9f07ad
+size 53525995

tacotron.toml

@@ -19,7 +19,7 @@ PAD_TOKEN = 0
 TEST_DATA_SIZE = 1024
 
 # model
-RR = 2 # reduction factor
+RR = 1 # reduction factor
 MAX_RR=2
 ATTN_BIAS = 0.0 # control how slow the attention moves forward
 SIGMOID_NOISE = 2.0