You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Thumbnail

End-To-End TEXT-2-ASMR with Transformers

This repository contains pretrained text2asmr model files, audio files and training+inference notebooks.

Dataset Details

This unique dataset is tailored for training and deploying text-to-speech (TTS) systems specifically focused on ASMR (Autonomous Sensory Meridian Response) content. It includes a comprehensive collection of pretrained model files, audio files and training code suitable for TTS applications.

Dataset Description

Inside this dataset, you shall find zipped folders as is follows:

  1. wavs_original: original wav files as it was converted from the original video
  2. wavs: original wav files broken into 1 minute chunks
  3. transcripts_original: transribed scripts of the original wav files
  4. transcripts: transribed scripts of the files in wav folder
  5. models: text to spectrogram model trained on Glow-TTS
  6. ljspeech: alignment files and respective checkpoint models (text to phoneme)
  7. transformer_tts_data.ljspeech: trained checkpoint models and other files

And the following files:

  1. Glow-TTS.ipynb: Training and inference code for GlowTTS models
  2. TransformerTTS.ipynb: Training and inference code for Transformer models
  3. VITS_TTS.ipynb: Optional code for training VITS models; follows the same format as GlowTTS
  4. metadata_original.csv: ljspeech formatted transcriptions of wav_original folder; ready for TTS training
  5. metadata.csv: ljspeech formatted transcriptions of wav folder; ready for TTS training

Latest Update: End-To-End TEXT-2-ASMR with Diffusion

Based on the paper, E3 TTS: EASY END-TO-END DIFFUSION-BASED TEXT TO SPEECH

(Yuan Gao, Nobuyuki Morioka, Yu Zhang, Nanxin Chen) Google

A text-to-asmr UNet-Diffusion model differing slightly from the framework mentioned in the paper was trained on the same audio-transcript paired dataset for 1000DDPM and 10 epochs.

Model metrics:

  1. General Loss: 0.000134
  2. MSE Loss: 0.000027
  3. RMSE Loss: 0.000217
  4. MAE Loss: 0.000018
  • Curated by: Alosh Denny, Anish S
  • Language(s) (NLP): English
  • License: MIT

Dataset Sources

Youtube: Rebeccas ASMR, Nanou ASMR, Gibi ASMR, Cherie Lorraine ASMR, etc.

Uses

The dataset can be used to train text2spec2mel, text2wav, and/or other end-to-end text-to-speech models.

Direct Use

Pretrained models can be tested out with the TransformerTTS notebook and the Glow-TTS notebook.

Dataset Card Authors

Alosh Denny, Anish S

Dataset Card Contact

[email protected]

Downloads last month
208