February 12, 2019

681 words 4 mins read



Deep neural networks for voice conversion (voice style transfer) in Tensorflow

repo name andabi/deep-voice-conversion
repo link https://github.com/andabi/deep-voice-conversion
language Python
size (curr.) 172486 kB
stars (curr.) 3033
created 2017-10-16
license MIT License

Voice Conversion with Non-Parallel Data

Subtitle: Speaking like Kate Winslet

Authors: Dabi Ahn(andabi412@gmail.com), Kyubyong Park(kbpark.linguist@gmail.com)




What if you could imitate a famous celebrity’s voice or sing like a famous singer? This project started with a goal to convert someone’s voice to a specific target voice. So called, it’s voice style transfer. We worked on this project that aims to convert someone’s voice to a famous English actress Kate Winslet’s voice. We implemented a deep neural networks to achieve that and more than 2 hours of audio book sentences read by Kate Winslet are used as a dataset.

Model Architecture

This is a many-to-one voice conversion system. The main significance of this work is that we could generate a target speaker’s utterances without parallel data like <source’s wav, target’s wav>, <wav, text> or <wav, phone>, but only waveforms of the target speaker. (To make these parallel datasets needs a lot of effort.) All we need in this project is a number of waveforms of the target speaker’s utterances and only a small set of <wav, phone> pairs from a number of anonymous speakers.

The model architecture consists of two modules:

  1. Net1(phoneme classification) classify someone’s utterances to one of phoneme classes at every timestep.
    • Phonemes are speaker-independent while waveforms are speaker-dependent.
  2. Net2(speech synthesis) synthesize speeches of the target speaker from the phones.

We applied CBHG(1-D convolution bank + highway network + bidirectional GRU) modules that are mentioned in Tacotron. CBHG is known to be good for capturing features from sequential data.

Net1 is a classifier.

  • Process: wav -> spectrogram -> mfccs -> phoneme dist.
  • Net1 classifies spectrogram to phonemes that consists of 60 English phonemes at every timestep.
    • For each timestep, the input is log magnitude spectrogram and the target is phoneme dist.
  • Objective function is cross entropy loss.
  • TIMIT dataset used.
    • contains 630 speakers' utterances and corresponding phones that speaks similar sentences.
  • Over 70% test accuracy

Net2 is a synthesizer.

Net2 contains Net1 as a sub-network.

  • Process: net1(wav -> spectrogram -> mfccs -> phoneme dist.) -> spectrogram -> wav
  • Net2 synthesizes the target speaker’s speeches.
    • The input/target is a set of target speaker’s utterances.
  • Since Net1 is already trained in previous step, the remaining part only should be trained in this step.
  • Loss is reconstruction error between input and target. (L2 distance)
  • Datasets
    • Target1(anonymous female): Arctic dataset (public)
    • Target2(Kate Winslet): over 2 hours of audio book sentences read by her (private)
  • Griffin-Lim reconstruction when reverting wav from spectrogram.



  • python 2.7
  • tensorflow >= 1.1
  • numpy >= 1.11.1
  • librosa == 0.5.1


  • sample rate: 16,000Hz
  • window length: 25ms
  • hop length: 5ms


  • Train phase: Net1 and Net2 should be trained sequentially.
    • Train1(training Net1)
      • Run train1.py to train and eval1.py to test.
    • Train2(training Net2)
      • Run train2.py to train and eval2.py to test.
        • Train2 should be trained after Train1 is done!
  • Convert phase: feed forward to Net2
    • Run convert.py to get result samples.
    • Check Tensorboard’s audio tab to listen the samples.
    • Take a look at phoneme dist. visualization on Tensorboard’s image tab.
      • x-axis represents phoneme classes and y-axis represents timesteps
      • the first class of x-axis means silence.

Tips (Lessons We’ve learned from this project)

  • Window length and hop length have to be small enough to be able to fit in only a phoneme.
  • Obviously, sample rate, window length and hop length should be same in both Net1 and Net2.
  • Before ISTFT(spectrogram to waveforms), emphasizing on the predicted spectrogram by applying power of 1.0~2.0 is helpful for removing noisy sound.
  • It seems that to apply temperature to softmax in Net1 is not so meaningful.
  • IMHO, the accuracy of Net1(phoneme classification) does not need to be so perfect.
    • Net2 can reach to near optimal when Net1 accuracy is correct to some extent.


comments powered by Disqus