Uberduck Written Tutorial
  • ‼️BEFORE YOU ASK ANYTHING, READ THIS FIRST!
  • 😊Introduction
  • 🗒️Plans and Updates
  • Discord Server Usage
    • Uberduck's Discord server walkthrough 100%
  • Requirements
    • 🤔Which Sources Should I Start?
    • 🛠️Requirements
  • Gathering Dialogs
    • Intro
      • Audacity Tutorial (WIN/MAC/Linux Users)
      • RX7/8 Tutorial (WIN/MAC Users)
      • FL Studio Tutorial (WIN/MAC Users) (WIP)
      • Vegas Pro Tutorial (Windows Only)
      • FFmpeg/FFBatch Tutorial (WIN/MAC/Linux Only)
      • Web version (Any platform)
  • Transcribing and fixing audios
    • Transcribing Dialogues
    • Transcribing Dialogues (Web Version)
    • Templates
    • Vocal Remover Colab Page Guide
  • Dataset sharing
    • Compressing Dataset
  • 🌮Installing Tacotron 2 Locally
    • 🐍Python Clean Install
    • 💻Miniconda Setup
  • Training time
    • Easy or Hard?
    • Training Audio With Tacotron 2 Colab Page
      • Simplified Training Page
      • Advanced Training Page
    • Setting Up for Local Training (WIP)
  • Testing Audio model
    • Testing Voice Model
    • Testing Voice Model (CPU)
  • Troubleshoot and bugs
    • Troubleshooting
  • Extra Stuff
    • Talknet usage
  • Wrapping up
    • The end
Powered by GitBook
On this page
  • List of errors
  • Wavs are not in mono and the sample rate is incorrect
  • Wavs are corrupted.
  • Model is broken
  • Oops, I have some empty lines on my text file.
  • I got this lady's default voice, which is that's not my model. What to do?
  • I ran in to CUDA's Out of memory.
  • I can't hear anything on synthesis page.
  • The wavs are in non-data
  • Feedback

Was this helpful?

  1. Troubleshoot and bugs

Troubleshooting

PreviousTesting Voice Model (CPU)NextTalknet usage

Last updated 3 years ago

Was this helpful?

Some people may encounter some issues and errors on colab page. Here's a list of things of errors that you have:

List of errors

Wavs are not in mono and the sample rate is incorrect

RuntimeError: shape '[1, 1, 51652]' is invalid for input of size
ValueError: wavs/6.wav 192000 SR doesn't match target 22050 SR

Well, that error meaning you have stereo and not mono and your sample rate is not 22050. You need to convert your audio files into mono and sample rate 22050. There should be a batch conversion tutorial if you have already done that with audios.

Wavs are corrupted.

ValueError: Incomplete wav chunk.

One of your wavs is probably corrupted.

Model is broken

PytorchStreamReader failed reading zip archive: failed finding central directory

Apparently, your model seems to be corrupted. What you wanna do is to retrain from scratch (by deleting your model on your google drive and start over)

Oops, I have some empty lines on my text file.

IndexError: list index out of range

Your text document has some empty lines to it, remove some empty lines and you should be good to go.

I got this lady's default voice, which is that's not my model. What to do?

You have probably low audios, it needs more audios to train tacotron. Go up to 50 or 100 and do the training process. Or probably tacotron bug. It must be iterations go up. Which it should be always 0.

I ran in to CUDA's Out of memory.

RuntimeError: CUDA out of memory. Tried to allocate 20.00 MiB 
(GPU 0; 14.76 GiB total capacity; 13.12 GiB already allocated; 
15.75 MiB free; 13.71 GiB reserved in total by PyTorch)

You need to lower your batch size, try lowering it somewhere 17 or 15 and you should be good to go. Remember, don't put batch size too much.

I can't hear anything on synthesis page.

Probably because on Uberduck's synthesis colab page (Waveglow) doesn't have supported GPU. Especially, let's say it doesn't show which GPU been used. Remember, you need at least

  • P100

  • V100

  • T4

And not K80 or P4. If you get these GPU that is not on list then do Factory reset via, Runtime --> Factory Reset Runtime. And remember it should be pop up on your side, or else if it doesn't pop up, you won't just factory reset.

The wavs are in non-data

WavFileWarning: Chunk (non-data) not understood, skipping it.

Most likely, some of a wavs have some metadata, it needs to be cleared away.

The easiest way to do is using Foobar2000, just drag all your wav files, select all, right click and select properties and you should see metadata, you need to clear all of it, once it's clear, generate mels again.

Feedback

If you have some lists of errors and answers on that, I'll be happy to put them on my guide.

Just ping me on uberduck's discord server (i disabled server dm's for myself so I won't get random dm's like that, especially I'm not accepting any discord friends requests unless I know you or met you before)

FFmpeg/FFBatch Tutorial (WIN/MAC/Linux Only)