keonlee9420 / portaspeech

PyTorch Implementation of PortaSpeech: Portable and High-Quality Generative Text-to-Speech

PortaSpeech - PyTorch Implementation

PyTorch Implementation of PortaSpeech: Portable and High-Quality Generative Text-to-Speech.

Audio Samples

Audio samples are available at /demo.

Model Size

Module	Normal	Small	Normal (paper)	Small (paper)
Total	24M	7.6M	21.8M	6.7M
LinguisticEncoder	3.7M	1.4M	-	-
VariationalGenerator	11M	2.8M	-	-
FlowPostNet	9.3M	3.4M	-	-

Quickstart

DATASET refers to the names of datasets such as LJSpeech in the following documents.

Dependencies

You can install the Python dependencies with

pip3 install -r requirements.txt

Also, Dockerfile is provided for Docker users.

Inference

You have to download the pretrained models and put them in output/ckpt/DATASET/.

For a single-speaker TTS, run

python3 synthesize.py --text "YOUR_DESIRED_TEXT" --restore_step RESTORE_STEP --mode single --dataset DATASET

The generated utterances will be put in output/result/.

Batch Inference

Batch inference is also supported, try

python3 synthesize.py --source preprocessed_data/DATASET/val.txt --restore_step RESTORE_STEP --mode batch --dataset DATASET

to synthesize all utterances in preprocessed_data/DATASET/val.txt.

Controllability

The speaking rate of the synthesized utterances can be controlled by specifying the desired duration ratios. For example, one can increase the speaking rate by 20 by

python3 synthesize.py --text "YOUR_DESIRED_TEXT" --restore_step RESTORE_STEP --mode single --dataset DATASET --duration_control 0.8

Please note that the controllability is originated from FastSpeech2 and not a vital interest of PortaSpeech.

Training

Datasets

The supported datasets are

• LJSpeech: a single-speaker English dataset consists of 13100 short audio clips of a female speaker reading passages from 7 non-fiction books, approximately 24 hours in total.

Preprocessing

Run

python3 prepare_align.py --dataset DATASET

for some preparations.

For the forced alignment, Montreal Forced Aligner (MFA) is used to obtain the alignments between the utterances and the phoneme sequences. Pre-extracted alignments for the datasets are provided here. You have to unzip the files in preprocessed_data/DATASET/TextGrid/. Alternately, you can run the aligner by yourself.

After that, run the preprocessing script by

python3 preprocess.py --dataset DATASET

Training

Train your model with

python3 train.py --dataset DATASET

Useful options:

• To use Automatic Mixed Precision, append --use_amp argument to the above command.
• The trainer assumes single-node multi-GPU training. To use specific GPUs, specify CUDA_VISIBLE_DEVICES=<GPU_IDs> at the beginning of the above command.

TensorBoard

Use

tensorboard --logdir output/log

to serve TensorBoard on your localhost. The loss curves, synthesized mel-spectrograms, and audios are shown.

Normal Model

Small Model Loss

Notes

• For vocoder, HiFi-GAN and MelGAN are supported.
• No ReLU activation and LayerNorm in VariationalGenerator to avoid mashed output.
• Speed ​​up the convergence of word-to-phoneme alignment in LinguisticEncoder by dividing long words into subwords and sorting the dataset by mel-spectrogram frame length.
• There are two kinds of helper loss to improve word-to-phoneme alignment: "ctc" and "dga". You can toggle them as follows:

  # In the train.yaml
  aligner:
      helper_type: "dga" # ["dga", "ctc", "none"]

  • "dga": Diagonal Guided Attention (DGA) Loss
  • "ctc": Connectionist Temporal Classification (CTC) Loss with forward-sum algorithm
  • If you set "none", no helper loss will be applied during training.
  • The alignments comparision of three methods ("dga", "ctc", and "none" from top to bottom):
  • The default setting is "dga". Although "ctc" makes the strongest alignment, the output quality and the accuracy are worse than "dga".
  • But still, there is a room for the improvement of output quality. The audio quality and the alingment (accuracy) seem to be a trade-off.
• Will be extended to a multi-speaker TTS.

Citation

Please cite this repository by the "Cite this repository" of About section (top right of the main page).

References

• jaywalnut310's VITS
• jaywalnut310's Glow-TTS
• keonlee9420's VAENAR-TTS
• keonlee9420's Comprehensive-Transformer-TTS (CTC Loss)
• keonlee9420's Comprehensive-Tacotron2 (DGA Loss)

portaspeech's Issues

Is the preprocess alignment necessary?

for multispeaker dataset train tts, is this preprocess alignment necessary? This provided preprocess script support Chinese?

About def get_mask_from_lengths(lengths, max_len=None):

Hi@keonlee9420, Thank You very much!
def get_mask_from_lengths(lengths, max_len=None):
batch_size = lengths.shape[0]
if max_len is None:
max_len = torch.max(lengths).item()

ids = torch.arange(0, max_len).unsqueeze(
    0).expand(batch_size, -1).to(lengths.device)
mask = ids >= lengths.unsqueeze(1).expand(-1, max_len)

return ~mask

In PortaSpeech, the return is ~mask, while in DiffGAN-TTS it is mask. I want to know the difference between them!

A run Problem(LJSpeech)

File "preprocess.py", line 20, in
preprocessor.build_from_path()
File "D:\UW-Detection\PortaSpeech\preprocessor\preprocessor.py", line 129, in build_from_path
n_frames += n
UnboundLocalError: local variable 'n' referenced before assignment
when I use NATSpeech show this problem (BiaoBei dataset)
when I use LJSpeech dataset and this code show this problem
I use global but cannot ......

Training data required

Hello,
Thanks for the code. Do you know if I can fine tune the model with 30 mins of data?

Is that possible train Chinese?

Is that possible train Chinese?

What happens with word_encoder if some words are out of vocab ?

Missing pretrained models

Hi, the link to the pretrained models seems to lead to an empty directory. Could you fix?

The meaning of inputs[11:] in model.loss.py

HI@[keonlee9420]，I cannot understand the meaning of inputs[11:] in model.loss.py

def forward(self, inputs, predictions, step):
(
mel_targets,
*_,
) = inputs[11:]
Thank you very much!

What's the key difference between this repo and the official repo?

Hi @keonlee9420 , thanks for your code and work. But I want to know what's the key difference between this repo and the official repo?

missing keys

Traceback (most recent call last):
File "synthesize.py", line 153, in
model = get_model(args, configs, device, train=False)
File "/content/PortaSpeech/utils/model.py", line 21, in get_model
model.load_state_dict(ckpt["model"])
File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1407, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for PortaSpeech:
Missing key(s) in state_dict: "linguistic_encoder.phoneme_encoder.attn_layers.3.emb_rel_k", "linguistic_encoder.phoneme_encoder.attn_layers.3.emb_rel_v", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_q.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_q.bias", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_k.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_k.bias", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_v.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_v.bias", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_o.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_o.bias", "linguistic_encoder.phoneme_encoder.norm_layers_1.3.gamma", "linguistic_encoder.phoneme_encoder.norm_layers_1.3.beta", "linguistic_encoder.phoneme_encoder.ffn_layers.3.conv.weight", "linguistic_encoder.phoneme_encoder.ffn_layers.3.conv.bias", "linguistic_encoder.phoneme_encoder.norm_layers_2.3.gamma", "linguistic_encoder.phoneme_encoder.norm_layers_2.3.beta", "linguistic_encoder.word_encoder.attn_layers.3.emb_rel_k", "linguistic_encoder.word_encoder.attn_layers.3.emb_rel_v", "linguistic_encoder.word_encoder.attn_layers.3.conv_q.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_q.bias", "linguistic_encoder.word_encoder.attn_layers.3.conv_k.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_k.bias", "linguistic_encoder.word_encoder.attn_layers.3.conv_v.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_v.bias", "linguistic_encoder.word_encoder.attn_layers.3.conv_o.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_o.bias", "linguistic_encoder.word_encoder.norm_layers_1.3.gamma", "linguistic_encoder.word_encoder.norm_layers_1.3.beta", "linguistic_encoder.word_encoder.ffn_layers.3.conv.weight", "linguistic_encoder.word_encoder.ffn_layers.3.conv.bias", "linguistic_encoder.word_encoder.norm_layers_2.3.gamma", "linguistic_encoder.word_encoder.norm_layers_2.3.beta", "variational_generator.flow.flows.0.enc.in_layers.3.bias", "variational_generator.flow.flows.0.enc.in_layers.3.weight_g", "variational_generator.flow.flows.0.enc.in_layers.3.weight_v", "variational_generator.flow.flows.0.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.0.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.0.enc.res_skip_layers.3.weight_v", "variational_generator.flow.flows.2.enc.in_layers.3.bias", "variational_generator.flow.flows.2.enc.in_layers.3.weight_g", "variational_generator.flow.flows.2.enc.in_layers.3.weight_v", "variational_generator.flow.flows.2.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.2.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.2.enc.res_skip_layers.3.weight_v", "variational_generator.flow.flows.4.enc.in_layers.3.bias", "variational_generator.flow.flows.4.enc.in_layers.3.weight_g", "variational_generator.flow.flows.4.enc.in_layers.3.weight_v", "variational_generator.flow.flows.4.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.4.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.4.enc.res_skip_layers.3.weight_v", "variational_generator.flow.flows.6.enc.in_layers.3.bias", "variational_generator.flow.flows.6.enc.in_layers.3.weight_g", "variational_generator.flow.flows.6.enc.in_layers.3.weight_v", "variational_generator.flow.flows.6.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.6.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.6.enc.res_skip_layers.3.weight_v", "variational_generator.dec_wn.in_layers.3.bias", "variational_generator.dec_wn.in_layers.3.weight_g", "variational_generator.dec_wn.in_layers.3.weight_v", "variational_generator.dec_wn.res_skip_layers.3.bias", "variational_generator.dec_wn.res_skip_layers.3.weight_g", "variational_generator.dec_wn.res_skip_layers.3.weight_v", "postnet.flows.24.logs", "postnet.flows.24.bias", "postnet.flows.25.weight", "postnet.flows.26.start.bias", "postnet.flows.26.start.weight_g", "postnet.flows.26.start.weight_v", "postnet.flows.26.end.weight", "postnet.flows.26.end.bias", "postnet.flows.26.cond_layer.bias", "postnet.flows.26.cond_layer.weight_g", "postnet.flows.26.cond_layer.weight_v", "postnet.flows.26.wn.in_layers.0.bias", "postnet.flows.26.wn.in_layers.0.weight_g", "postnet.flows.26.wn.in_layers.0.weight_v", "postnet.flows.26.wn.in_layers.1.bias", "postnet.flows.26.wn.in_layers.1.weight_g", "postnet.flows.26.wn.in_layers.1.weight_v", "postnet.flows.26.wn.in_layers.2.bias", "postnet.flows.26.wn.in_layers.2.weight_g", "postnet.flows.26.wn.in_layers.2.weight_v", "postnet.flows.26.wn.res_skip_layers.0.bias", "postnet.flows.26.wn.res_skip_layers.0.weight_g", "postnet.flows.26.wn.res_skip_layers.0.weight_v", "postnet.flows.26.wn.res_skip_layers.1.bias", "postnet.flows.26.wn.res_skip_layers.1.weight_g", "postnet.flows.26.wn.res_skip_layers.1.weight_v", "postnet.flows.26.wn.res_skip_layers.2.bias", "postnet.flows.26.wn.res_skip_layers.2.weight_g", "postnet.flows.26.wn.res_skip_layers.2.weight_v", "postnet.flows.27.logs", "postnet.flows.27.bias", "postnet.flows.28.weight", "postnet.flows.29.start.bias", "postnet.flows.29.start.weight_g", "postnet.flows.29.start.weight_v", "postnet.flows.29.end.weight", "postnet.flows.29.end.bias", "postnet.flows.29.cond_layer.bias", "postnet.flows.29.cond_layer.weight_g", "postnet.flows.29.cond_layer.weight_v", "postnet.flows.29.wn.in_layers.0.bias", "postnet.flows.29.wn.in_layers.0.weight_g", "postnet.flows.29.wn.in_layers.0.weight_v", "postnet.flows.29.wn.in_layers.1.bias", "postnet.flows.29.wn.in_layers.1.weight_g", "postnet.flows.29.wn.in_layers.1.weight_v", "postnet.flows.29.wn.in_layers.2.bias", "postnet.flows.29.wn.in_layers.2.weight_g", "postnet.flows.29.wn.in_layers.2.weight_v", "postnet.flows.29.wn.res_skip_layers.0.bias", "postnet.flows.29.wn.res_skip_layers.0.weight_g", "postnet.flows.29.wn.res_skip_layers.0.weight_v", "postnet.flows.29.wn.res_skip_layers.1.bias", "postnet.flows.29.wn.res_skip_layers.1.weight_g", "postnet.flows.29.wn.res_skip_layers.1.weight_v", "postnet.flows.29.wn.res_skip_layers.2.bias", "postnet.flows.29.wn.res_skip_layers.2.weight_g", "postnet.flows.29.wn.res_skip_layers.2.weight_v", "postnet.flows.30.logs", "postnet.flows.30.bias", "postnet.flows.31.weight", "postnet.flows.32.start.bias", "postnet.flows.32.start.weight_g", "postnet.flows.32.start.weight_v", "postnet.flows.32.end.weight", "postnet.flows.32.end.bias", "postnet.flows.32.cond_layer.bias", "postnet.flows.32.cond_layer.weight_g", "postnet.flows.32.cond_layer.weight_v", "postnet.flows.32.wn.in_layers.0.bias", "postnet.flows.32.wn.in_layers.0.weight_g", "postnet.flows.32.wn.in_layers.0.weight_v", "postnet.flows.32.wn.in_layers.1.bias", "postnet.flows.32.wn.in_layers.1.weight_g", "postnet.flows.32.wn.in_layers.1.weight_v", "postnet.flows.32.wn.in_layers.2.bias", "postnet.flows.32.wn.in_layers.2.weight_g", "postnet.flows.32.wn.in_layers.2.weight_v", "postnet.flows.32.wn.res_skip_layers.0.bias", "postnet.flows.32.wn.res_skip_layers.0.weight_g", "postnet.flows.32.wn.res_skip_layers.0.weight_v", "postnet.flows.32.wn.res_skip_layers.1.bias", "postnet.flows.32.wn.res_skip_layers.1.weight_g", "postnet.flows.32.wn.res_skip_layers.1.weight_v", "postnet.flows.32.wn.res_skip_layers.2.bias", "postnet.flows.32.wn.res_skip_layers.2.weight_g", "postnet.flows.32.wn.res_skip_layers.2.weight_v", "postnet.flows.33.logs", "postnet.flows.33.bias", "postnet.flows.34.weight", "postnet.flows.35.start.bias", "postnet.flows.35.start.weight_g", "postnet.flows.35.start.weight_v", "postnet.flows.35.end.weight", "postnet.flows.35.end.bias", "postnet.flows.35.cond_layer.bias", "postnet.flows.35.cond_layer.weight_g", "postnet.flows.35.cond_layer.weight_v", "postnet.flows.35.wn.in_layers.0.bias", "postnet.flows.35.wn.in_layers.0.weight_g", "postnet.flows.35.wn.in_layers.0.weight_v", "postnet.flows.35.wn.in_layers.1.bias", "postnet.flows.35.wn.in_layers.1.weight_g", "postnet.flows.35.wn.in_layers.1.weight_v", "postnet.flows.35.wn.in_layers.2.bias", "postnet.flows.35.wn.in_layers.2.weight_g", "postnet.flows.35.wn.in_layers.2.weight_v", "postnet.flows.35.wn.res_skip_layers.0.bias", "postnet.flows.35.wn.res_skip_layers.0.weight_g", "postnet.flows.35.wn.res_skip_layers.0.weight_v", "postnet.flows.35.wn.res_skip_layers.1.bias", "postnet.flows.35.wn.res_skip_layers.1.weight_g", "postnet.flows.35.wn.res_skip_layers.1.weight_v", "postnet.flows.35.wn.res_skip_layers.2.bias", "postnet.flows.35.wn.res_skip_layers.2.weight_g", "postnet.flows.35.wn.res_skip_layers.2.weight_v".
size mismatch for linguistic_encoder.abs_position_enc: copying a param with shape torch.Size([1, 1001, 128]) from checkpoint, the shape in current model is torch.Size([1, 1001, 192]).

Noise at the end of the speech

hi, In the ljspeech dataset speech demo you gave, there is noise at the end of the speech. I have the same problem in the Chinese and English datasets during training. The quality of speech generation is not very good. Any suggestions

无法训练

请问下，现在的程序是不是训练不了呢？ 在text的处理上，代码采用的是字符作为单元，而不是空格分开的音素为单元。发现跟phones_per_word对应不上。就是 len(phones)!=sum(phones_per_word)，我分析到的原因是phones采用的字符为单元。

A questions about the output of Phoneme Encoding

Hi@keonlee9420, after the linguistic encoder is implemented, the text is input to the character embedding layer, and the output value contains Nan. How to solve this problem?

Multi-speaker TTS

Dear sir,

First of all, I really appriciate your contribution in this amazing repo! However, it would be perfect if you can add the feature of multi-speaker TTS here. I can see the spker_emb was not used now. Do I know when can you consider this and opmimize the ability of this impressive model!

Thanks,

Max

small model

saw that the normal model for LJSpeech was released thanks, will the small model also be released?

RuntimeError: The size of tensor a (256) must match the size of tensor b (45) at non-singleton dimension 2

Hi @keonlee9420, the a(256) is encoder_hidden and the b(45) is the first max_time of src_seq. in the word_level_pooling. I very want to know how to solve the problem. Thank you very much!

I notice that the b(45) is variable.

Weird sound from the beginning of the sentence "hello"

Hi, thanks for your contribution in TTS, and it's such a great work !!
It's seems perfect in most of the sentence while trying python3 synthesize.py --text "MY_SENTENCE" --restore_step 125000 --mode single --dataset LJSpeech, but when I tried the sentence with "hello" in the front, the sound of "hello" became long and weird. Here is the mel-spetrogram of "Hello, glad to see you."

And you can observe a large area on the left is represent the word "hello" clearly.
I've tried to check your training data in preprocessed_data/LJSpeech/train.txt, and I couldn't find the word "hello" in that.

Is this problem caused by the quantity of the phoneme of the word merely or I just do something wrong or something else?
Anything would help, thank you.

Speed in CPU

Hi, thank you very much for you work and share.
In the paper, the proposed method have been compared with many methods in MOS, parameter size, as well as the speed. While you compute the RTF with GPU, did you compared the RTF / speed when running in CPU?

RuntimeError: Found dtype Long but expected Float

File "train.py", line 122, in main
model_update(model, step, G_loss, optG_fs2)
File "train.py", line 77, in model_update
loss = (loss / grad_acc_step).backward()
File "C:\Users\12604\Anaconda3\envs\pytorch\lib\site-packages\torch\tensor.py", line 221, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "C:\Users\12604\Anaconda3\envs\pytorch\lib\site-packages\torch\autograd_init_.py", line 132, in backward
allow_unreachable=True) # allow_unreachable flag
RuntimeError: Found dtype Long but expected Float

Hi@keonlee9420. This problem occurs when the loss function is back-propagating, how can I solve it?
This is the dtype of loss

I hear echo & background noise on the DEMO wavs

This can happen because of blurring of mel-spectrogram(in most cases start of the sound should have sharp edge), which can happen because bad attention. I could recommend to try using Diagonal guided attention (DGA) during the training(https://arxiv.org/abs/1710.08969)

Inference issue

Basically, I tried to run it in the Google Colab

1st cell

%cd /content/
!git clone https://github.com/keonlee9420/PortaSpeech
%cd /content/PortaSpeech/
!pip install -r /content/PortaSpeech/requirements.txt

2nd

id_big = '1VTotGmE42a19bevwgQ9mhPkXzQvKzl8q'
id_small = '1Y0IGlc4zJ7XN5sh4aPWLTeQ80D9ZhfbB'

!mkdir /content/PortaSpeech/output/
!mkdir /content/PortaSpeech/output/ckpt/
!mkdir /content/PortaSpeech/output/ckpt/DATASET/
%cd /content/PortaSpeech/output/ckpt/DATASET/
!gdown --id $id_big 
!gdown --id $id_small 
%cd /content/PortaSpeech

3rd

%cd /content/PortaSpeech
!python3 synthesize.py --text "Moved to Site-19 1993. Origin is as of yet unknown. It is constructed from concrete and rebar with traces of Krylon brand spray paint." \
                        --restore_step 125000 --mode single --dataset DATASET

and this is what I've got:

/content/PortaSpeech
[nltk_data] Downloading package averaged_perceptron_tagger to
[nltk_data]     /root/nltk_data...
[nltk_data]   Unzipping taggers/averaged_perceptron_tagger.zip.
[nltk_data] Downloading package cmudict to /root/nltk_data...
[nltk_data]   Unzipping corpora/cmudict.zip.
2021-10-26 10:57:51.803863: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
Traceback (most recent call last):
  File "synthesize.py", line 138, in <module>
    args.dataset)
  File "/content/PortaSpeech/utils/tools.py", line 19, in get_configs_of
    os.path.join(config_dir, "preprocess.yaml"), "r"), Loader=yaml.FullLoader)
FileNotFoundError: [Errno 2] No such file or directory: './config/DATASET/preprocess.yaml'

What this 'preprocess.yaml' is exactly?

Can this project run in windows？

无法训练

请问下，现在的程序是不是训练不了呢？ 在text的处理上，代码采用的是字符作为单元，而不是空格分开的音素为单元。发现跟phones_per_word对应不上。就是 len(phones)!=sum(phones_per_word)，我分析到的原因是phones采用的字符为单元。

Is PortaSpeech a better choice than FastSpeech2 or DiffSpeech?

From your experience, how are the effects of these models ranked.

After training the AISHELL3 dataset, The synthesized sound is all electric current sound

Hi @keonlee9420, after training the AISHELL3 dataset, The synthesized sound is all electric current sound. I want to know how to happed.

why valid loss is raising?

why valid loss is raising?

any example vioce samples?

small(320000.pth.tar) weights incompatibility

`2022-11-11 22:31:08.004017: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cudart64_110.dll

Device of PortaSpeech: cpu
Traceback (most recent call last):
File "synthesize.py", line 153, in
model = get_model(args, configs, device, train=False)
File "D:\projects\PortaSpeech\utils\model.py", line 21, in get_model
model.load_state_dict(ckpt["model"])
File "C:\ProgramData\Miniconda3\envs\tts_env\lib\site-packages\torch\nn\modules\module.py", line 1223, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for PortaSpeech:
Missing key(s) in state_dict: "linguistic_encoder.phoneme_encoder.attn_layers.3.emb_rel_k", "linguistic_encoder.phoneme_encoder.attn_layers.3.emb_rel_v", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_q.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_q.bias", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_k.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_k.bias", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_v.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_v.bias", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_o.weight", "linguistic_encoder.phoneme_encoder.attn_layers.3.conv_o.bias", "linguistic_encoder.phoneme_encoder.norm_layers_1.3.gamma", "linguistic_encoder.phoneme_encoder.norm_layers_1.3.beta", "linguistic_encoder.phoneme_encoder.ffn_layers.3.conv.weight", "linguistic_encoder.phoneme_encoder.ffn_layers.3.conv.bias", "linguistic_encoder.phoneme_encoder.norm_layers_2.3.gamma", "linguistic_encoder.phoneme_encoder.norm_layers_2.3.beta", "linguistic_encoder.word_encoder.attn_layers.3.emb_rel_k", "linguistic_encoder.word_encoder.attn_layers.3.emb_rel_v", "linguistic_encoder.word_encoder.attn_layers.3.conv_q.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_q.bias", "linguistic_encoder.word_encoder.attn_layers.3.conv_k.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_k.bias", "linguistic_encoder.word_encoder.attn_layers.3.conv_v.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_v.bias", "linguistic_encoder.word_encoder.attn_layers.3.conv_o.weight", "linguistic_encoder.word_encoder.attn_layers.3.conv_o.bias", "linguistic_encoder.word_encoder.norm_layers_1.3.gamma", "linguistic_encoder.word_encoder.norm_layers_1.3.beta", "linguistic_encoder.word_encoder.ffn_layers.3.conv.weight", "linguistic_encoder.word_encoder.ffn_layers.3.conv.bias", "linguistic_encoder.word_encoder.norm_layers_2.3.gamma", "linguistic_encoder.word_encoder.norm_layers_2.3.beta", "variational_generator.flow.flows.0.enc.in_layers.3.bias", "variational_generator.flow.flows.0.enc.in_layers.3.weight_g", "variational_generator.flow.flows.0.enc.in_layers.3.weight_v", "variational_generator.flow.flows.0.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.0.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.0.enc.res_skip_layers.3.weight_v", "variational_generator.flow.flows.2.enc.in_layers.3.bias", "variational_generator.flow.flows.2.enc.in_layers.3.weight_g", "variational_generator.flow.flows.2.enc.in_layers.3.weight_v", "variational_generator.flow.flows.2.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.2.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.2.enc.res_skip_layers.3.weight_v", "variational_generator.flow.flows.4.enc.in_layers.3.bias", "variational_generator.flow.flows.4.enc.in_layers.3.weight_g", "variational_generator.flow.flows.4.enc.in_layers.3.weight_v", "variational_generator.flow.flows.4.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.4.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.4.enc.res_skip_layers.3.weight_v", "variational_generator.flow.flows.6.enc.in_layers.3.bias", "variational_generator.flow.flows.6.enc.in_layers.3.weight_g", "variational_generator.flow.flows.6.enc.in_layers.3.weight_v", "variational_generator.flow.flows.6.enc.res_skip_layers.3.bias", "variational_generator.flow.flows.6.enc.res_skip_layers.3.weight_g", "variational_generator.flow.flows.6.enc.res_skip_layers.3.weight_v", "variational_generator.dec_wn.in_layers.3.bias", "variational_generator.dec_wn.in_layers.3.weight_g", "variational_generator.dec_wn.in_layers.3.weight_v", "variational_generator.dec_wn.res_skip_layers.3.bias", "variational_generator.dec_wn.res_skip_layers.3.weight_g", "variational_generator.dec_wn.res_skip_layers.3.weight_v", "postnet.flows.24.logs", "postnet.flows.24.bias", "postnet.flows.25.weight", "postnet.flows.26.start.bias", "postnet.flows.26.start.weight_g", "postnet.flows.26.start.weight_v", "postnet.flows.26.end.weight", "postnet.flows.26.end.bias", "postnet.flows.26.cond_layer.bias", "postnet.flows.26.cond_layer.weight_g", "postnet.flows.26.cond_layer.weight_v", "postnet.flows.26.wn.in_layers.0.bias", "postnet.flows.26.wn.in_layers.0.weight_g", "postnet.flows.26.wn.in_layers.0.weight_v", "postnet.flows.26.wn.in_layers.1.bias", "postnet.flows.26.wn.in_layers.1.weight_g", "postnet.flows.26.wn.in_layers.1.weight_v", "postnet.flows.26.wn.in_layers.2.bias", "postnet.flows.26.wn.in_layers.2.weight_g", "postnet.flows.26.wn.in_layers.2.weight_v", "postnet.flows.26.wn.res_skip_layers.0.bias", "postnet.flows.26.wn.res_skip_layers.0.weight_g", "postnet.flows.26.wn.res_skip_layers.0.weight_v", "postnet.flows.26.wn.res_skip_layers.1.bias", "postnet.flows.26.wn.res_skip_layers.1.weight_g", "postnet.flows.26.wn.res_skip_layers.1.weight_v", "postnet.flows.26.wn.res_skip_layers.2.bias", "postnet.flows.26.wn.res_skip_layers.2.weight_g", "postnet.flows.26.wn.res_skip_layers.2.weight_v", "postnet.flows.27.logs", "postnet.flows.27.bias", "postnet.flows.28.weight", "postnet.flows.29.start.bias", "postnet.flows.29.start.weight_g", "postnet.flows.29.start.weight_v", "postnet.flows.29.end.weight", "postnet.flows.29.end.bias", "postnet.flows.29.cond_layer.bias", "postnet.flows.29.cond_layer.weight_g", "postnet.flows.29.cond_layer.weight_v", "postnet.flows.29.wn.in_layers.0.bias", "postnet.flows.29.wn.in_layers.0.weight_g", "postnet.flows.29.wn.in_layers.0.weight_v", "postnet.flows.29.wn.in_layers.1.bias", "postnet.flows.29.wn.in_layers.1.weight_g", "postnet.flows.29.wn.in_layers.1.weight_v", "postnet.flows.29.wn.in_layers.2.bias", "postnet.flows.29.wn.in_layers.2.weight_g", "postnet.flows.29.wn.in_layers.2.weight_v", "postnet.flows.29.wn.res_skip_layers.0.bias", "postnet.flows.29.wn.res_skip_layers.0.weight_g", "postnet.flows.29.wn.res_skip_layers.0.weight_v", "postnet.flows.29.wn.res_skip_layers.1.bias", "postnet.flows.29.wn.res_skip_layers.1.weight_g", "postnet.flows.29.wn.res_skip_layers.1.weight_v", "postnet.flows.29.wn.res_skip_layers.2.bias", "postnet.flows.29.wn.res_skip_layers.2.weight_g", "postnet.flows.29.wn.res_skip_layers.2.weight_v", "postnet.flows.30.logs", "postnet.flows.30.bias", "postnet.flows.31.weight", "postnet.flows.32.start.bias", "postnet.flows.32.start.weight_g", "postnet.flows.32.start.weight_v", "postnet.flows.32.end.weight", "postnet.flows.32.end.bias", "postnet.flows.32.cond_layer.bias", "postnet.flows.32.cond_layer.weight_g", "postnet.flows.32.cond_layer.weight_v", "postnet.flows.32.wn.in_layers.0.bias", "postnet.flows.32.wn.in_layers.0.weight_g", "postnet.flows.32.wn.in_layers.0.weight_v", "postnet.flows.32.wn.in_layers.1.bias", "postnet.flows.32.wn.in_layers.1.weight_g", "postnet.flows.32.wn.in_layers.1.weight_v", "postnet.flows.32.wn.in_layers.2.bias", "postnet.flows.32.wn.in_layers.2.weight_g", "postnet.flows.32.wn.in_layers.2.weight_v", "postnet.flows.32.wn.res_skip_layers.0.bias", "postnet.flows.32.wn.res_skip_layers.0.weight_g", "postnet.flows.32.wn.res_skip_layers.0.weight_v", "postnet.flows.32.wn.res_skip_layers.1.bias", "postnet.flows.32.wn.res_skip_layers.1.weight_g", "postnet.flows.32.wn.res_skip_layers.1.weight_v", "postnet.flows.32.wn.res_skip_layers.2.bias", "postnet.flows.32.wn.res_skip_layers.2.weight_g", "postnet.flows.32.wn.res_skip_layers.2.weight_v", "postnet.flows.33.logs", "postnet.flows.33.bias", "postnet.flows.34.weight", "postnet.flows.35.start.bias", "postnet.flows.35.start.weight_g", "postnet.flows.35.start.weight_v", "postnet.flows.35.end.weight", "postnet.flows.35.end.bias", "postnet.flows.35.cond_layer.bias", "postnet.flows.35.cond_layer.weight_g", "postnet.flows.35.cond_layer.weight_v", "postnet.flows.35.wn.in_layers.0.bias", "postnet.flows.35.wn.in_layers.0.weight_g", "postnet.flows.35.wn.in_layers.0.weight_v", "postnet.flows.35.wn.in_layers.1.bias", "postnet.flows.35.wn.in_layers.1.weight_g", "postnet.flows.35.wn.in_layers.1.weight_v", "postnet.flows.35.wn.in_layers.2.bias", "postnet.flows.35.wn.in_layers.2.weight_g", "postnet.flows.35.wn.in_layers.2.weight_v", "postnet.flows.35.wn.res_skip_layers.0.bias", "postnet.flows.35.wn.res_skip_layers.0.weight_g", "postnet.flows.35.wn.res_skip_layers.0.weight_v", "postnet.flows.35.wn.res_skip_layers.1.bias", "postnet.flows.35.wn.res_skip_layers.1.weight_g", "postnet.flows.35.wn.res_skip_layers.1.weight_v", "postnet.flows.35.wn.res_skip_layers.2.bias", "postnet.flows.35.wn.res_skip_layers.2.weight_g", "postnet.flows.35.wn.res_skip_layers.2.weight_v".
size mismatch for linguistic_encoder.abs_position_enc: copying a param with shape torch.Size([1, 1001, 128]) from checkpoint, the shape in current model is torch.Size([1, 1001, 192]).
size mismatch for linguistic_encoder.kv_position_enc: copying a param with shape torch.Size([1, 1001, 128]) from checkpoint, the shape in current model is torch.Size([1, 1001, 192]).
size mismatch for linguistic_encoder.q_position_enc: copying a param with shape torch.Size([1, 1001, 128]) from checkpoint, the shape in current model is torch.Size([1, 1001, 192]).
size mismatch for linguistic_encoder.src_emb.weight: copying a param with shape torch.Size([361, 128]) from checkpoint, the shape in current model is torch.Size([361, 192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.emb_rel_k: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.emb_rel_v: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_q.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_q.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_k.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_k.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_v.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_v.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_o.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.0.conv_o.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.emb_rel_k: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.emb_rel_v: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_q.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_q.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_k.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_k.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_v.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_v.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_o.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.1.conv_o.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.emb_rel_k: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.emb_rel_v: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_q.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_q.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_k.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_k.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_v.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_v.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_o.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.phoneme_encoder.attn_layers.2.conv_o.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_1.0.gamma: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_1.0.beta: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_1.1.gamma: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_1.1.beta: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_1.2.gamma: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_1.2.beta: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.ffn_layers.0.conv.weight: copying a param with shape torch.Size([128, 128, 3]) from checkpoint, the shape in current model is torch.Size([192, 192, 5]).
size mismatch for linguistic_encoder.phoneme_encoder.ffn_layers.0.conv.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.ffn_layers.1.conv.weight: copying a param with shape torch.Size([128, 128, 3]) from checkpoint, the shape in current model is torch.Size([192, 192, 5]).
size mismatch for linguistic_encoder.phoneme_encoder.ffn_layers.1.conv.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.ffn_layers.2.conv.weight: copying a param with shape torch.Size([128, 128, 3]) from checkpoint, the shape in current model is torch.Size([192, 192, 5]).
size mismatch for linguistic_encoder.phoneme_encoder.ffn_layers.2.conv.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_2.0.gamma: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_2.0.beta: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_2.1.gamma: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_2.1.beta: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_2.2.gamma: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.phoneme_encoder.norm_layers_2.2.beta: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.emb_rel_k: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.emb_rel_v: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_q.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_q.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_k.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_k.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_v.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_v.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_o.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.0.conv_o.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.emb_rel_k: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.emb_rel_v: copying a param with shape torch.Size([1, 9, 64]) from checkpoint, the shape in current model is torch.Size([1, 9, 96]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_q.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_q.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_k.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_k.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_v.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_v.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_o.weight: copying a param with shape torch.Size([128, 128, 1]) from checkpoint, the shape in current model is torch.Size([192, 192, 1]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.1.conv_o.bias: copying a param with shape torch.Size([128]) from checkpoint, the shape in current model is torch.Size([192]).
size mismatch for linguistic_encoder.word_encoder.attn_layers.2.emb_rel_k: copying a param with shape torch.Size([1, 9, 64]) from...`

Who can share the pre-trained model which is the AISHELL3

Who can share the pre-trained model which is the AISHELL3