> 🇵🇱 Polski (Mekatron) by @User Training: Synthesis: Hi-Fi Gan synthesis (CPU): HiFi Gan Training: __Pipeline__ Training: Synthesis:

> 🇧🇷 Português (PT-BR) by @User Training: Synthesis: __**Pipeline**__ Training: Synthesis:

Many community members have found these notebooks to be valuable for training voice models and making datasets. Many thanks to the people who authored the original versions of the notebooks, such as Cookie, The PPP, and so on, as those notebooks are not authored or maintained by Uberduck. > 🌮 Tacotron 2 🤖 To train a legacy Tacotron 2 model, use this notebook: https://colab.research.google.com/drive/1CIPTTj94EocZe2w5-zCDC3G44K9OaFal To test/use a legacy Tacotron 2 model—especially before submitting it to Uberduck, use this notebook: https://colab.research.google.com/drive/1NVA3ndxhYWsKn-zwh3NnzMMgoVdJ5xUx To train a pipeline Tacotron 2 model, use this notebook: https://colab.research.google.com/drive/1jF-Otw2_ssEcus4ISaIZu3QDmtifUvyY — Kaggle Simpson: https://www.kaggle.com/coldfir4/uberduck-tacotron-2-multispeaker-gsts-training To test/use a pipeline Tacotron 2 model—especially before submitting it to Uberduck, use this notebook: https://colab.research.google.com/drive/1g9W1stWS6RdeLT9PT5vIgXk_C19fnSx9 > 🦜 TalkNet 2 🥅 To train a TalkNet 2 model, use this notebook: https://colab.research.google.com/github/justinjohn0306/TalkNET-colab/blob/main/TalkNet_Training.ipynb To test/use a TalkNet 2 model—especially before submitting it to Uberduck, use this notebook: https://colab.research.google.com/drive/1xnuizJ4_URPc05USwvl84boAy8Q49fJO?usp=sharing or—preferably—this notebook: https://colab.research.google.com/drive/1aj6Jk8cpRw7SsN3JSYCv57CrR6s0gYPB Note: Even though we refer to it as "TalkNet", the AI architecture is actually TalkNet 2. > ✍️ Transcription ⌨️ Automatic Transcribing (compatible with GPU and CPU): https://colab.research.google.com/drive/1i0UXcvxXRW1yp9p4cymoZg_ZqsFBFyTB

ﾠ > Other Useful Notebooks  22kHz, mono, 16 bit .wav-inator: https://colab.research.google.com/drive/100uSy4yDgr118RfuSojVumI23nM4OBkY?usp=sharing#scrollTo=IXbqIuGG8aiW — This converts incorrectly formatted audio files into correctly formatted audio files so that a model can be trained with them.  Sample Rate Checker: https://colab.research.google.com/drive/18rKAA6Sg-xE4D2abwQsxuRWYkWcwrskt — This can be used to detect if there are any incorrectly formatted audio files.  To isolate speech and (kind of) remove background noise from audio, use this notebook: https://colab.research.google.com/github/NaJeongMo/Colaboratory-Notebook-for-Ultimate-Vocal-Remover/blob/main/Vocal%20Remover%205_arch.ipynb ⬆️ For this notebook, it is recommended to use this configuration:

pretrained_model: Vocal_HP_4BAND_3090_arch-124m.pth

window_size: 320

parameter: auto-detect

aggressiveness: 0.5

TTA: on

deepExtraction: on

isVocal: on

download: on or off, it's your choice

export_as_mp3: on or off, it's your choice

(Alternatively, you can use this site, which isn't a notebook: https://www.lalal.ai)  HiFi-GAN vocoder trainers for Tacotron 2: For legacy TT2: https://colab.research.google.com/drive/1ume3953K2K-EdNL90vNqPNSWM1KRuwqp For pipeline TT2: https://colab.research.google.com/drive/1SKu2xRJy5q1wzuP5CSO8dJ-Nf-UIKz0K

ﾠ > Extra Notebooks  GST multispeaker example notebook: https://colab.research.google.com/drive/1--mNMSNv5_OkqVLcYg_Znb4WX-1Br_zd Old quotation:

Interested if this works for you all.  The training time is pretty slow with the 3 speakers, and would be even slower with enough data for the GST to pick up on the interesting stuff, but hopefully this is useful?  It also uses ARPAbet, which we've noticed is pretty helpful.

 Backup version of the TalkNet output generator notebook that you can use in case the other ones are broken: https://colab.research.google.com/github/justinjohn0306/TalkNET-colab/blob/main/Controllable_TalkNet.ipynb

Auto Transcribing using OpenAI's Whisper: