shifathosn
/
artcraft
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

working on adapting the waveglow/inference to sound non-noisy

This commit is contained in:
Brandon Thomas 2021-06-02 22:17:27 -04:00
parent 077a4239d8
commit 7744f0e324
3 changed files with 16 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
.env
View File

@ -41,4 +41,6 @@ W2L_DOWNLOAD_SCRIPT_NAME=vocodes_process_upload.py
TTS_CODE_DIRECTORY=/home/bt/dev/storyteller/storyteller-web/models/tacotron2
TTS_INFERENCE_SCRIPT_NAME=vocodes_inference.py
TTS_VOCODER_MODEL_FILENAME=waveglow_256channels_universal_v5.pt
#TTS_VOCODER_MODEL_FILENAME=waveglow_256channels_universal_v5.pt # THIS SOUNDS HORRID
#TTS_VOCODER_MODEL_FILENAME=waveglow_256channels_v4_uberduck.pt # THIS IS TOO SMALL! 300MB vs 1.2 GB!
TTS_VOCODER_MODEL_FILENAME=waveglow_256channels_v4.pt

2
debug/curl/seed.py
View File

@ -114,6 +114,8 @@ upload_tts_model('https://drive.google.com/file/d/1atFn4vm7mjkufXNK9TZS6J8WzISrA
upload_tts_model('https://drive.google.com/file/d/1-rlJpi2jAsagn-UCStcCc_ITNycXgn2F/view?usp=sharing', 'Jorgen', session_other)
upload_tts_model('https://drive.google.com/file/d/1togTNOEKnaarqWWrbZLSwIbZIZ8IT_qw/view?usp=sharing', 'edsheeransinging', session_nanashi)
upload_tts_model('https://drive.google.com/file/d/1OAiu0cZDyUJMemyLc0Btqw7zEQgiN81D/view?usp=sharing', 'Wakko Warner v2 ST21', session_nanashi)
#upload_tts_model('', '', session_other)
#upload_tts_model('', '', session_other)
#upload_tts_model('', '', session_other)

22
models/tacotron2/vocodes_inference.py
View File

@ -20,14 +20,16 @@ args = parser.parse_args()
import sys
sys.path.append('waveglow/')
import numpy as np
import torch
import shutil
import json
from model import Tacotron2
import numpy as np
import torch
from hparams import create_hparams
from train import load_model
from model import Tacotron2
from layers import TacotronSTFT
from audio_processing import griffin_lim
from text import text_to_sequence
from denoiser import Denoiser
@ -41,20 +43,18 @@ print('========================================', flush=True)
# NB(bt, 2021-05-31): Trying to get everything on the same device
#torch.device("cuda" if torch.cuda.is_available() else "cpu")
hparams = create_hparams()
hparams.sampling_rate = 22050
hparams.sampling_rate = 22050 # Don't change this
hparams.max_decoder_steps = 1000 # How long the audio will be before it cuts off (1000 is about 11 seconds)
hparams.gate_threshold = 0.1 # Model must be 90% sure the clip is over before ending generation (the higher this number is, the more likely that the AI will keep generating until it reaches the Max Decoder Steps)
# Fix bug running on CPU?
#torch.set_default_dtype(torch.float16)
# NB(bt, 2021-05-31): Perhaps this is blocking waveglow? It complains of the wrong tensor type.
#torch.set_default_tensor_type('torch.DoubleTensor') # 2021-05-31 commented out
#torch.set_default_tensor_type('torch.HalfTensor') # 2021-05-31 added
# Load synthesizer
hparams = create_hparams()
hparams.sampling_rate = 22050 # Don't change this
hparams.max_decoder_steps = 1000 # How long the audio will be before it cuts off (1000 is about 11 seconds)
hparams.gate_threshold = 0.1 # Model must be 90% sure the clip is over before ending generation (the higher this number is, the more likely that the AI will keep generating until it reaches the Max Decoder Steps)
# Load synthesizer
checkpoint_path = args.synthesizer_checkpoint_path
#model = load_model(hparams)