Global Rhythm Style Transfer Without Text Transcriptions

Kaizhi Qian

Last update: Dec 30, 2022

Related tags

Overview

Global Prosody Style Transfer Without Text Transcriptions

This repository provides a PyTorch implementation of AutoPST, which enables unsupervised global prosody conversion without text transcriptions.

This is a short video that explains the main concepts of our work. If you find this work useful and use it in your research, please consider citing our paper.

@InProceedings{pmlr-v139-qian21b,
  title = 	 {Global Prosody Style Transfer Without Text Transcriptions},
  author =       {Qian, Kaizhi and Zhang, Yang and Chang, Shiyu and Xiong, Jinjun and Gan, Chuang and Cox, David and Hasegawa-Johnson, Mark},
  booktitle = 	 {Proceedings of the 38th International Conference on Machine Learning},
  pages = 	 {8650--8660},
  year = 	 {2021},
  editor = 	 {Meila, Marina and Zhang, Tong},
  volume = 	 {139},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {18--24 Jul},
  publisher =    {PMLR},
  url = 	 {http://proceedings.mlr.press/v139/qian21b.html}
}

Audio Demo

The audio demo for AutoPST can be found here

Dependencies

Python 3.6
Numpy
Scipy
PyTorch == v1.6.0
librosa
pysptk
soundfile
wavenet_vocoder pip install wavenet_vocoder==0.1.1 for more information, please refer to https://github.com/r9y9/wavenet_vocoder

To Run Demo

Download pre-trained models to assets

Download the same WaveNet vocoder model as in AutoVC to assets

Please refer to AutoVC if you have any problems with the vocoder part, because they share the same vocoder scripts.

Run demo.ipynb

To Train

Download training data to assets. The provided training data is very small for code verification purpose only. Please use the scripts to prepare your own data for training.

Prepare training data: python prepare_train_data.py
Train 1st Stage: python main_1.py
Train 2nd Stage: python main_2.py

Final Words

This project is part of an ongoing research. We hope this repo is useful for your research. If you need any help or have any suggestions on improving the framework, please raise an issue and we will do our best to get back to you as soon as possible.

Comments

How to test AutoPST in onother languages?

I have not success to test final section of vocoder code in autopst. In conda envionment, all dependencies have installed. Error something like"from synthessis cannot import build_model" allways shows. I want to ask you, do I need to train in AutoVC speakers, cause I want to use for my own sounds, in another language, not English?

just want to clone voice of one recording to onother. for different speakers. Do these recordings need to be same lenght and same sentences spoken to be compared during training?

BTW, I have RTX3060, and this card not supported by version of 1.6.0 of pytorch. I have installed fist onmt python package, than pytorch 1.7.0 with Cuda 11 Thank you

opened by zelenooki87 6
Missing basic execution with different set of speakers.

Hi there, I am trying to follow the code with my own dataset and could run Main_1.py and main_2.py to get xxx-A.ckpt and xxx-B.ckpt files. Now I am not able to understand to run the demo file to prepare specific speakers dictionary to create and convert. Any help is appreciated with a little more direction to follow the steps.

opened by sainishalini 4
Unable to reproduce results

@auspicious3000 We tried reproducing results using your codebase and the dataset found here https://datashare.ed.ac.uk/handle/10283/3443 (that you use) but unfortunately, we were unable to. The outputs that we have so far are extremely noisy (even if source and target speakers are the same). Could you please share a working code with us that might help us reproduce the results? I would greatly appreciate your inputs!

opened by avanitanna 1
Issue with stop prediction for longer utterances.

Hi @auspicious3000,

First, thanks for releasing this repository! I've been trying to compare AutoPST to some upcoming work but I'm having an issue with the stop token prediction when converting utterances longer than 1 or 2 seconds. I noticed that you clipped some of the VCTK files for your demo page (and in the test dictionary you provided) so that they're much shorter. How did you use the test utterances in your evaluations? Do you have any recommendations so that I can make as fair a comparison as possible.

Thanks, Benjamin

opened by bshall 1
SpeechSplit actually better than AutoPST for seen speakers?

Hello, I'm reading D.6. SPEECHSPLIT Baseline in the paper.

Am I understanding this correctly that SpeechSplit performs better for rythm transfer when the doing conversation to a seen speaker ?

opened by skol101 1
How can we generate test_vctk.meta?

Hi @auspicious3000, can you please provide me with the code to generate test_vctk.meta? The format of train_vctk.meta (which is generated using prepare_train_data.py) is very different from test. I would like to test your code on new data.

opened by avanitanna 5
How to solve SEA model problem

Hello. I have referred to your paper. Based on your experiment, I conducted experiment on accent transformation using English accent data from different countries. But the result is very unsatisfactory, I can't even hear the transformed voice clearly.

I think there may be a problem in the process of training SEA model. But I don't know exactly where the problem is.

The images show my code for training SEA. Could you help me with this issue? What is the best approach to solve this?

opened by sera920 0
KeyError when run prepare_train_data.py

Hi, I got an error like this, when run prepare_train_data.py Is spk2emb has vctk16-train-wav key?

vctk16-train-wav Traceback (most recent call last): File "prepare_train_data.py", line 52, in submeta.append(spk2emb[subdir]) KeyError: 'vctk16-train-wav'

opened by Jwaminju 2
ModuleNotFoundError: No module named 'onmt'

Hi, I run into an error about onmt

ModuleNotFoundError Traceback (most recent call last) in 5 import torch.nn.functional as F 6 from collections import OrderedDict ----> 7 from onmt.utils.misc import sequence_mask 8 from model_autopst import Generator_2 as Predictor 9 from hparams_autopst import hparams

ModuleNotFoundError: No module named 'onmt'

but I see folder onmt_modules exists

then I install onmt(pip install onmt) and notice it's installing torch 1.3.0 although the requirements say that PyTorch == v1.6.0

Could you help me with this issue? What is the best approach to solve this?

opened by stalevna 1
Inference with new input audio

Hi and thank you for this amazing project!

I was trying to create a notebook in colab that would allow me to input an audio file, then select the speaker and produce an output accordingly.

Here the code, it works but I am missing the part on how to change speaker timbre. Do you have any tips on that?

Thanks a lot in advance!

opened by shoegazerstella 4

Global Rhythm Style Transfer Without Text Transcriptions

Related tags

Overview

Global Prosody Style Transfer Without Text Transcriptions

Audio Demo

Dependencies

To Run Demo

To Train

Final Words

Comments

Owner

Kaizhi Qian

This code extends the neural style transfer image processing technique to video by generating smooth transitions between several reference style images

Rhythm-Finder is a unsupervised ML driven python powered web-application that can find the songs that suits you.

Text to speech is a process to convert any text into voice. Text to speech project takes words on digital devices and convert them into audio. Here I have used Google-text-to-speech library popularly known as gTTS library to convert text file to .mp3 file. Hope you like my project!

Unet-TTS: Improving Unseen Speaker and Style Transfer in One-shot Voice Cloning

Code for the paper "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer"

Kashgari is a production-level NLP Transfer learning framework built on top of tf.keras for text-labeling and text-classification, includes Word2Vec, BERT, and GPT2 Language Embedding.

Code for the paper "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer"

Kashgari is a production-level NLP Transfer learning framework built on top of tf.keras for text-labeling and text-classification, includes Word2Vec, BERT, and GPT2 Language Embedding.

An easy to use, user-friendly and efficient code for extracting OpenAI CLIP (Global/Grid) features from image and text respectively.

Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis (SV2TTS)

Ptorch NLU, a Chinese text classification and sequence annotation toolkit, supports multi class and multi label classification tasks of Chinese long text and short text, and supports sequence annotation tasks such as Chinese named entity recognition, part of speech tagging and word segmentation.

Text-Summarization-using-NLP - Text Summarization using NLP to fetch BBC News Article and summarize its text and also it includes custom article Summarization

Extract city and country mentions from Text like GeoText without regex, but FlashText, a Aho-Corasick implementation.

Global Tracking Transformers, CVPR 2022

Scikit-learn style model finetuning for NLP

Scikit-learn style model finetuning for NLP

:house_with_garden: Fast & easy transfer learning for NLP. Harvesting language models for the industry. Focus on Question Answering.

:house_with_garden: Fast & easy transfer learning for NLP. Harvesting language models for the industry. Focus on Question Answering.

Code for our ACL 2021 paper - ConSERT: A Contrastive Framework for Self-Supervised Sentence Representation Transfer