# Applio Colab

Last update: Mar 10, 2024


image
#

# Introduction ‎

  • Applio is a fork of Mangio developed by the IA Hispano team.

  • It's liked for its great UI & lots of extra features, such as TTS (with RVC models too), plugins, automatic model upload, customizable theme & more.

  • Because of its user-friendly experience & active development, it's considered to be one of the best forks.

  • As this cloud version is hosted in Google Colab, remember that you have a runtime of 4 hours.

# Pros & Cons

✔️ PROS
CONS
  • Very complete
  • Has an active development
  • TTS features
  • Automatic model upload
  • Has Mangio-Crepe
  • User-friendly UI
  • A little slower compared w/ forks
  • More unstable
  • Usage limit for free users

#

# Setting Up

#
  1. Access the Colab space here. Then log in to your Google account.

#
  1. Execute the Install Applio cell. This will take around 2 minutes.

    image

  • It'll finish when you see a tick symbol on the left.

    image


#
  1. If you are going to train models, upload your dataset to your Google Drive storage & run the Extra cell.

    image

  • To save time, unfold it & cancel the custom pretrain download, if you aren't going to use them.

    image


#
  1. Grant the permissions to Google Drive.

    image

#
  1. Execute Start Applio.

    image

  • Then open the public URL.

    image


#

# Inference

#

# 1. Upload voice model.

  • Go to the Download tab.
    You have two ways of uploading it: through its link or manually inputting its files.

    1. Go to the Download tab & paste the link in the Model Link bar.
      It must be from Hugging Face or Google Drive.

      image
    2. Press Download Model.
    1. Below in Drop files, press the upload box & input the model's .PTH.

      image
    2. Then input the .INDEX.

# 2. Select voice model.

  1. Return to the Inference tab & click Refresh on the right.

    image

  2. Select the model in the Voice Model & Index File dropdown.

    image

#

# 3. Input vocals.

  • With Applio you can convert audios individually or in batches:

    1. Press the upload box & input your audio.

      image

    2. Then select it in the dropdown below.

      image
    1. Go to the Batch tab.

      image

    2. Go to the file explorer in Colab. Go to drive, right-click the folder containing the audios & click Copy Path.

    3. Paste the path in the Input Folder bar.

    • In Output Folder you can define the path folder for the results.

    • Ensure the paths don't contain spaces/special characters.

# 4. Modify settings. (optional)

  • Unfold Advanced Settings if you wish to modify the inference settings for better results.

    image


#

# 5. Convert.

  1. Click Convert at the bottom to process the audio.

  2. Once it's done, you can hear the results in the Export Audio box below.
    To download it, press the download symbol on its right.

    image


#
#

# Training

#
#
# a. Model Name
#
  • Go to the Train tab. Input a name for your model in Model Name.
    Don't include spaces/special characters.

    image


#
# b. Dataset Path
#
  1. Upload your dataset to your GD storage if you haven't already.
  2. In Colab click the folder on the left ( ) & click the reload button.

    image

    (For mobile users: tap the three lines on the top left & Show file browser)
  3. Open drive, localize your dataset, right-click it & click Copy path.

    image

  4. Then paste it on the Dataset Path bar.

    image

#
# c. Sampling Rate
#
  • Select your dataset's sample rate. If you don't know the amount, click here.

    image


#
# d. Preprocess Dataset
#
  • Ensure RVC Version is set as V2 & click Preprocess Dataset.

    It'll finish when the output box says preprocessed successfully.

    image
#
# a. Pitch extraction algorithm
#
  • Select the algorithm you want. Use either Crepe or RMVPE, as the rest are outdated.

    image

#
# b. Hop Length (optional)
#
  • If you chose Crepe, you can modify its hop length.

    image

#
# c. Extract Features
#
  • Press Extract Features.
    It'll finish when it says extracted successfully.

    image
#
# a. Batch Size
#
  • If you are a newbie, use 8. But in case your dataset is short (around 2 minutes or less), use 4.

    image

#
# b. Save Every Epoch
#
  • Frequency of the saving checkpoints, based on the epochs.

  • If you are a newbie, simply leave it at 15.

    image

  • E.g: with a value of 10, they will be saved after the epoch 10, 20, 30, etc.


#
# c. Total Epoch
#
  • Input the total amount of epochs (training cycles) for the model.

  • But since we'll use TensorBoard, use an arbitrarily large value like 1000

    image


#
# d. Generate Index
#
  • Click Generate Index. This will create the model's .INDEX file.

#
# e. Start Training
#
  1. Tick Save Only Latest

    image

  2. Press Start Training below to begin the training process.

    image


#
# f. Monitor training
#
  1. TB will be available in the Colab. Remember to monitor it, as well as the cell's logs just in case.

    The latter will show you errors if they happen, and information about the epochs & checkpoints.

    image

  2. If after around 2:30 hours of training you don't detect OT download the model of the lowest point, in case it's already OT, and the .INDEX.

  3. Then once your GPU runtime resets, begin the retraining procedure.


#
# a. Stop training
#
  • When you're very sure of overtraining, you can stop training by going to the Settings tab & press Restart Applio.

    image

  • Come back to the Colab & open the new public URL.


#
# b. Get the INDEX
#
  1. Open the file explorer, go to logs, and open the folder named as the model.
  2. Download the .INDEX named added_.

    image‎ ‎

#
# c. Get the PTH
#
  1. In said folder you'll also find all the checkpoints.

  2. Select the one closest to before the overtraining point, and move it to the new folder.

    • The checkpoints will be organized with this format: ModelName_Epoch.pth
      Example: arianagrande_60e.pth

      image

    • You can determine the Step number of the checkpoints by looking at its epoch number on the logs.

      image

  3. And that's all, have fun with your model. To test it, do a normal inference as usual.

#
  • In case the training finished but the model still needed training, you don't have to start from scratch.
  1. Simply enter the same settings & criteria that you had previously inserted. You don't have to do preprocess, extract feature or train the .INDEX again.
  2. You can change the save frequency or increase the Total Epoch amount, in case you didn't input enough before.
  3. If you're resuming from a new session, unfold the Extra cell in Colab & input the model name you assigned before.

    image
  • For this, the Auto Backup cell must've ran in the previous session.

    image
  1. Begin training again & remember to monitor TB as before.
#
#

# TTS

+ with any RVC model

#
  • Applio is also known for having one TTS tool by default, with plenty of voices to choose for.

  • You can also use it with RVC models & apply the inference settings if you wish.

  • Additionally, you can download the Eleven Labs TTS plugin.


#

# Instructions:

  1. Go to the TTS tab.

    image


#
  1. If you want to use an RVC model, download it, go to TTS, click Refresh & select it in Voice Model & Index File.

    image

  • To modify the inference settings or the output folder for the TTS/RVC audio, unfold Advanced Settings.

#
  1. In TTS Voices select the voice of your desired language, accent & gender.

    In Text to Synthesize input your text. Then click Convert.

    image

  • If you are using an RVC model, select a voice that matches the model the most, to guarantee great results.

#
  1. Once it's done, you'll be able to hear the result in the Export Audio box. To download it, click the download button on its right ( ).

    image


#
#

# Extra

#
  • Applio has an Extra menu, containing an audio analyzer, originally made by Ilaria.

  • Making it convenient for determining the sample rate of datasets when training models.

  • It also contains the model fusion tool, ideal for advanced users.


#

# Audio Analyzer:

  1. Go to the Extra tab & press the upload box to input your audio.

    image


#
  1. Once it's done uploading, click Get information about the audio.

#
  1. In Sampling rate you'll see the audio's full sample rate. Use said value for training.

    image

#
#
  • # Example:

    image

Here it reached 20 kHz. Doubling it gives 40kHz. Therefore the ideal target sample rate would be 40k


#
#

# Plugins

  • Plugins are components that you can add to Applio, that add new features & enhance your experience.

  • These are made by the public, and are free & easy to install.

  • You can find them on their GitHub page. More will be added in the future.


#

# Installation:

  1. Access their GitHub page & click on the name of the plugin you want.

    image


#
  1. Click on the ZIP file.

    image

  • Click on the download button on the right. This will download the ZIP file of the plugin.

    image


#
  1. Open Applio & head over to the Plugins tab. Press the upload box & upload the ZIP.

    image


#
  1. Go to the Settings tab & click Restart Applio at the bottom. Go back to the Colab & open the new public URL.

    Then you'll be able to see the plugin in the Plugins tab.

    image


#
#

# Troubleshooting

#
#
  • If it's lower than 32k: select 32k.
  • If it's 44.1k: select 40k.
  • If i'ts higher than 48k: select 48k.
#
  • This a phenomenon called artifacting. To fix it, read here.
#
#
  • Report your issue here.

#

# You have reached the end.

Report Issues