/asr_guided_tacotron

Use las to enhance the performance of tacotron, especially at the lack of the speaker labels.

Primary LanguagePythonMIT LicenseMIT

ASR Guided Tacotron

Use LAS model to enhance the performance of tacotron, especially at the lack of the speaker labels.

Prerequisites

  1. Python packages:

    • python 3.4 or higher
    • tensorflow r1.8 or higher
    • numpy
    • librosa
    • scipy
    • tqdm
    • matplotlib

  2. Clone this repository:

https://github.com/HappyBall/asr_guided_tacotron.git

Dataset

  1. multi-speaker dataset mixed from VCTK and LibriSpeech Download

Usage

Before training the whole model includes Tacotron model and LAS model, you need to pre-train both models respectively.

Also, you can download our pre-trained Tacotron models Here and LAS models Here to skip the pre-train procedures.

Pre-train LAS model

  1. Download the dataset and use transcript_training_las.txt as --train_data_name in the hyperparams.py.

  2. Set up the correct path of the dataset and other hyperparameters in hyperparams.py.

Run: python train_las.py --keep_train False

Parameter --keep_train determines either start a new training or continue training with the existed model which the path should be correctly set up in hyperparams.py.

Pre-train Tacotron model

  1. Download the dataset and use transcript_training_tacotron_seqlen50.txt as --train_data_name in the hyperparams.py.

  2. Set up the correct path of the dataset and other hyperparameters in hyperparams.py.

Run: python train_origin_tacotron.py --keep_train False

Train Tacotron model with the guidance of LAS

  1. Download the dataset and use transcript_training_tacotron_seqlen50.txt as --train_data_name in the hyperparams.py.

  2. Set up the correct paths of the pre-trained models and other hyperparameters in hyperparams.py.

Run: python train_tacotron.py --keep_train True

The program will automatically load the pre-trained models and start training with the guidance of ASR.

While training, only Tacotron model will be updated.

Synthesis

  1. Set up the correct path of the existed model and the reference audio file which you want to encode prosody from in hyperparams.py.

  2. Add English input sequences you want to synthesize into test_sentenses.txt.

Run: python synthesize.py

Hyperparameters of the hyperparams.py

--data: the path of the data directory which contains the wav files
--prepro_path: the path of the preprocessed data directory
--test_data: the path of the text file which contains input text sequences to synthesize speech
--ref_wavfile: the path of the reference audio file which you want to encode prosody from
--train_data_name: the transcription file name using for training
--taco_logdir: the path of the directory to save or load Tacotron models
--taco_logfile: the path of Tacotron training log file
--las_logdir: the path of the directory to save or load LAS models
--sampledir: the path of the directory to save speech files when synthesizing
--attention_mechanism: choose the type of attention mechanism (original or dot)
taco_consis_weight: determines how much would the attention consistency influence the loss function (decimal number from 0 to 1)
--n_iter: iteration number of Griffin Lim algorithm
--lr: initial learning rate

Files in this project

Folders:

las/: modules and networks of las model
tacotron/: modules and networks of tacotron model

Files:

data_load.py: data loader for training data and testing data
evaluate_las.py: calculate the character error rate (CER) of las model
graph.py: define model graph
hyperparams.py: set up training hyperparameters and directory for saving models
prepro.py: preprocess data
synthesize.py: synthesize speech conditioned on input text sequences in the test sentence file and the reference audio file
train_las.py: pre-train LAS models
train_origin_tacotron.py: pre-train Tacotron models without guidance of ASR
train_tacotron.py: train to improve existed Tacotron models with guidance of ASR