Tryed to run google collab, but there is an error. Please help.

How to implement it on local Computer?

Привет, у меня не видит dataset RU. LibriSpeech видит, а русский dataset не видит. Что делать? Спасибо

Hi. Thank you.
Is it possible to not train your own Tacotron2 and use pretrained model on russian language?
This for example https://github.com/alphacep/tn2-wg

first, thanks for such a complete pipeline.
second, would you integrate e.g. this repo for native russian support?

First of all, thank you for sharing the open-source of Multi-Tacotron-Voice-Cloning. I also just started learning about natural language processing programming. And I also started learning Python programming.
-I put the software in the directory: D: \ SV2TTS
-I put the dataset in the directory: D: \ Datasets, I have D: \ Datasets \ book and D: \ Datasets \ LibriSpeech

When using the code you provided, I had some training issues:

1. I have finished the steps

• Run python encoder_preprocess.py D: \ Datasets
  and the result is
  Arguments:
  datasets_root: D: \ Datasets
  out_dir: D: \ Datasets \ SV2TTS \ encoder
  datasets: ['preprocess_voxforge']
  skip_existing: False
  Done preprocessing book.

2. Run visdom
3. But I could not continue

• Run python encoder_train.py my_run D: \ Datasets
  because the notice appeared
  C: \ Users \ Admin \ anaconda3 \ envs \ [Test_Voice] \ lib \ site-packages \ umap \ spectral.py: 4: NumbaDeprecationWarning: No direct replacement for 'numba.targets' available. Visit https://gitter.im/numba/numba-dev to request help. Thanks!
  import numba.targets
  usage: encoder_train.py [-h] [--clean_data_root CLEAN_DATA_ROOT]
  [-m MODELS_DIR] [-v VIS_EVERY] [-u UMAP_EVERY]
  [-s SAVE_EVERY] [-b BACKUP_EVERY] [-f]
  [--visdom_server VISDOM_SERVER] [--no_visdom]
  run_id
  encoder_train.py: error: unrecognized arguments: D: \ Datasets

My question: How can I fix this problem?

Thanks again for your sharing!!!

Здравствуйте. У меня видео-карта GeForce GT 630M. Уж очень нужно запустить программу, но там у PyTorch минимум compute capability 3.0, а на моей видеокарте 2.1
Нужен именно интерфейс, для изучения графиков, на Colab только ввиде командной строки.
Если устанавливаю версию только CPU, то пишет, что не поддерживается.
Может быть есть способ отключить использование GPU и перейти на CPU?

Hi,

I'm slightly confused in G2P model. Let's suppose If I need to train a model which specifically translates from Russian to English (Only). Do I still need to add dictionary or train G2P model ?

Also, I'm not able to catch the significance of G2P model here. We have the synthesizer, which is already doing the same work.

Thanks!

Is it possible to use WaveGlow?
And not to train your own model but use pretrained?

Предположим нужно добиться хорошего звучания всего одного голоса(русский). Куда нажимать(инструкция для хомяков).

Hello, I am looking for a way So I can able to make a Japanese speaker speaks English, is it possible?

https://github.com/vlomme/Multi-Tacotron-Voice-Cloning/blob/master/g2p/train.py#L20

How can i add Arabic Language ??

Hi @vlomme,

Great work here, and thanks for open-sourcing it. I'm trying to understand how this works so that I can replicate it. I've gone through the code and don't see any language embedding, which I thought would be how you separate the speaker from the language.

Can you please explain how language-speaker independence is achieved?

Hi I am doing similar work like yours, my datasets is "En + Chineses".
I have tried the pretrained model offered by CorentinJ, and also finetune on the pretrained model, but i have not achieve good results till now. I am still training the encoder model now. And I wonder if you have some good results to share?

OSError: [WinError 127] Не найдена указанная процедура. Error loading "C:\Users\admin\PycharmProjects\pythonProject\venv\lib\site-packages\torch\lib\caffe2_detectron_ops_gpu.dll" or one of its dependencies.

Пробовал разные методы и подходы, в общем ничего не помогает... Даже не знаю что и делать.

P.S. tensorflow-gpu поддерживается только до python-3.7, всё что выше будет писать ошибку что не нашло нужной версии.

Может кто сталкивался?

Доброго времени суток!
Как начать С НУЛЯ обучать нейронку? (т.е не нужен pretrained model)

Спасибо больше за проект, я ищу какое-то быстрое решение для себя, которое могло бы делать voice style transfer (пародировать записанный голос по семплу), можно ли применить ваш синтезатор для этой задачи?

Hello,
I tried using the demo with the pre-trained network.
See the attached example. The output is just pure noise.
Is there anything wrong with the voice sample i provided as input?
I tried both male / female.

Thanks,
orig_16.zip

Thanks for work! Help me to train an encoder. How is it possible to add new custom voices to train datasets, or only fixed (like LibriSpeech: train-other-500, VoxCeleb1...) are available through the interface of commands:
python encoder_preprocess.py <datasets_root>
and
python encoder_train.py my_run <datasets_root>/SV2TTS/encoder

If possible, than how i should keep files, in root data directory or subfolders, in what formats? I tried to add my voice to subfolder but got an error like:

"Python encoder_preprocess.py data
Arguments:
datasets_root: data
out_dir: data/SV2TTS/encoder
datasets: ['preprocess_voxforge']
skip_existing: False

Preprocessing preprocess_voxforge
Couldn't find data/book, skipping this dataset"

I looked at the source and found that there are fixed funcs that preprocess different formats of train data (like preprocess22,preprocess44...) What do they mean? Maybe i should use one of them?
Thank you.

Hello,
Amazing work.
I am running inference using your models on 2080 gpu. your example is perfect. But when I give a new audio clip (in English) and make it say the same Russian sentence, the output audio isn't good. There's lot of noise, and cloning is not even of good quality.

My question is:

1. Can I use pretrained models(from this repo) to clone a new speaker, and make it speak Russian? or Should I train every thing(g2p, encoder, synthesizer, vocoder) on new speaker(assuming I obtain hours of this speaker's audio)? Please advise.

Thanks,
S

Запустил демо-версию, не работает.

Интересует твоё оценочное суждение, может сэмплы накидаешь, чтобы поиметь представление?
Я пробовал https://github.com/CorentinJ/Real-Time-Voice-Cloning и результатом не впечатлён, куча каких то шумов, на пару тонов выше речь.

Здравствуйте, экспериментировал с вашей моделью, но лишь некоторые записи дают хороший результат. На Хабре вы писали, что пробовали так же обучать модель только для русского языка и она работала лучше, у вас не осталось натренерованной модели? Если да, не могли бы вы поделиться ей?

I have a problem inserting additional text to be spoken into the toolbox. The additional lines cause the vocoder to crash with out of memory error. Trying with the original code from CorentinJ and code from here I found that activating g2p in toolbox / __ init__ caused this error.
Apparently the g2p binds the resources that are important for other neural networks and does not release them when it has finished its task.
Can you fix that somehow?

Thank you

File "...\g2p\train.py", line 4, in
from distance import levenshtein
ModuleNotFoundError: No module named 'distance'