Kyubyong / expressive_tacotron

Tensorflow Implementation of Expressive Tacotron

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

A TensorFlow Implementation of Expressive Tacotron

This project aims at implementing the paper, Towards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotron, to verify its concept. Most of the baseline codes are based on my previous Tacotron implementation.

Requirements

  • NumPy >= 1.11.1
  • TensorFlow >= 1.3
  • librosa
  • tqdm
  • matplotlib
  • scipy

Data

Because the paper used their internal data, I train the model on the LJ Speech Dataset

LJ Speech Dataset is recently widely used as a benchmark dataset in the TTS task because it is publicly available. It has 24 hours of reasonable quality samples.

Training

  • STEP 0. Download LJ Speech Dataset or prepare your own data.
  • STEP 1. Adjust hyper parameters in hyperparams.py. (If you want to do preprocessing, set prepro True`.
  • STEP 2. Run python train.py. (If you set prepro True, run python prepro.py first)
  • STEP 3. Run python eval.py regularly during training.

Sample Synthesis

I generate speech samples based on the same script as the one used for the original web demo. You can check it in test_sents.txt.

  • Run python synthesize.py and check the files in samples.

Samples

16 sample sentences in the first chapter of the original web demo are collected for sample synthesis. Two audio clips per sentence are used for prosody embedding--reference voice and base voice. Mostly, those two are different in terms of gender or region. The samples below look like the following:

  • 1a: the first reference audio
  • 1b: sample embedded with 1a's prosody
  • 1c: the second reference audio (base)
  • 1d: sample embedded with 1c's prosody

Check out the samples at each steps.

Analysis

  • Hearing the results of 130k steps, it's not clear if the model has learned the prosody.
  • It's clear that different reference audios cause different samples.
  • Some samples are worthy of note. For example, listen to the four audios of no.15. The stress of "right" part was obvious transferred.
  • Check out no.9, reference audios of which are sung. They are fun.

Notes

  • Because this repo focuses on investigating the concept of the paper, I did not follow some details of the paper.
  • The paper used phoneme inputs, whereas I stuck to graphemes.
  • Instead of the Bahdanau attention, the paper used the GMM attention.
  • The original audio samples were obtained from wavenet vocoder.
  • I'm still confused what the paper claims to be a prosody embedding can be isolated from the speaker.
  • For prosody embedding, the authors employed conv2d. Why not conv1d?
  • When the reference audio's text or sentence structure is totally different from the inference script, what happens?
  • If I have time, I'd like to implement their 2nd paper: Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis

April 2018, Kyubyong Park

About

Tensorflow Implementation of Expressive Tacotron


Languages

Language:Python 100.0%