Kahsolt / TransTacoS-RetuneGAN

A toy-like Text-to-Speech for Chinese/Mandarin synthesize, inspired by Tacotron & FastSpeech2 & RefineGAN.

TransTacoS-RetuneGAN

A lighter-weight (perhaps!) Text-to-Speech for Chinese/Mandarin synthesize, inspired by Tacotron & FastSpeech2 & RefineGAN.
It is also my shitty graduation design project, just a toy, so lower your expectations :)

---

Quick Start

setup

Since TransTacoS is implemented in tensorflow while RefineGAN in torch respectively, you could separate them by creating virtual envs, but they are likely not to conflict, thus you could try to put all these together:

• install tensorflow-gpu==1.14.0 tensorboard==1.14.0 following https://tensorflow.google.cn/install/pip
• install torch==1.8.0+cu1xx torchaudio==0.8.0 following https://pytorch.org/, where cu1xx is your cuda version
• run pip install -r requirements.txt for the rest dependencies

dataset

• download and unzip the open source dataset DataBaker
• other dataset requires user-defined preprocessor, please refer to transtacos/dataset/__skel__.py

train

• check path configs in all Makefile
• cd transtacos & make preprocess to prepare acoustic features (linear/mel/f0/c0/zcr)
• cd transtacos & make train to train TransTacoS
• cd retunegan & make finetune to train RetuneGAN using preprocessed linear spectrograms (rather than from raw wave)

deploy

• check port configs in all Makefile
• cd transtacos & make server to start TransTacoS headless HTTP server (default at port 5105)
• cd retunegan & make server to start RetuneGAN headless HTTP server (default at port 5104)
• python app.py to start the WebUI app (default at port 5103)
• point your browser to http://localhost:5103, now have a try!

Model Architecture

TransTacoS

What I actually did:

• TransTacoS :=
  • embed: self-designed G2P solution (syl4), employ prodosy marks as linguistic feature
    • for G2P solution, I tried char(seq), phoneme, pinyin (CV), CVC, VCV, CVVC ...
      • de facto, they merely influence mel_loss, but effect that how controllable the pronounciation is
      • so I design syl4 to split phoneme against tone (just small improvement)
    • I predict prodosy marks by simple CNN rather than RNN, it's not enough reasonable though..
  • encoder: modified from FastSpeech2
    • I use multi-head self-DotAttn with GFFW (gated-FFW) for backbone transform
    • f0/c0 feature is quantilized to embed, then fused in with cross-DotAttn
  • decoder: inherited from Tacotron
    • this RNN+LSA decoder is really complicated, I dare not to touch :(
  • posnet: self-designed simple MLP and grouped-Linear
    • I found simple Linear layer with growing depth leads to lower mel_loss than Conv1d
    • thus realy DO NOT understand why Tacotron2 even FastSpeech2 still use a postnet directly strech n_channels from 80 (n_mel) to 1025 (n_freq), this is surely tooooo hard to compensate information loss

Frankly speaking, TransTacoS didn't improve any thing profoundly from Tacotron, but I just found that shallower network leads to lower mel_loss, so maybe simple embed+decoder is already enough :(

Tips of ideas to try or failed:

• To predict f0/sp/ap features so that we can use WORLD vocoder
  • It seems sp is OK, because it resembles mel very much
  • but ap requires to be carefully normalized, and accurate f0 is even harder to predict
  • audio quality of WORLD sounds worse than Griffin-Lim at times :(
• To predict spectrograms' magnitude part together with phase part, so that we can directly vocode using pure istft
  • It should be hard to optimize losses on phase, especially on conjunction points between mel frames
  • but these guys claim that they did it: iSTFTNet
• To predict f0 and dyn so that vocoder might benefits
  • I don't know how to separate them from mel, because so far I must regularize decoder's output to be mel (for the sake of teacher force)
  • When I remove the mel loss, I found that the RNN decoder became lazy to learn, and the align model also not work
  • mel is even more quantized that linear, to abstract f0 and dyn from only mel seems not the reasonable
• To extract duration info from the soft-aligned alignment map, so that we can further train a non-autogressive FastSpeech2 to speed up inference
  • just like FastSpeech and DeepSinger did
  • but FastSpeech reported that extracted duration is not enough accurate, yet I could not fully understand and reproduce DeepSinger's DP algorithm for duration extraction

RetuneGAN

What I actually did:

• RetuneGAN :=
  • preprocess
    • extract reference wav using Griffin-Lim
    • extract u/v mask by hand-tuned zrc/c0 threshold (for Split-G only)
  • generators
    • UNet-G (encoder-decoder generator): modified from RefineGAN, we use the output of Griffin-Lim as reference wav, rather than an F0/C0-guided hand-crafted speech template
    • Split-G (split u/v generator): self-designed, inspired by Multi-Band MelGAN, but I found the generated quality is holy shit :(
      • ResStack borrowed from MelGAN
      • ResBlock modified from HiFiGAN
  • discriminators
    • MSD (multi scale discriminator): borrowed from MelGAN, I think it's good for plosive consonants
    • MPD (multi period discriminator): borrowed from HiFiGAN, I take it as a multiple MSDs' stack-up
    • MTD (multi stft discriminator): modified from UnivNet, it has two work modes depending on its input (MPSD seems better indeed ...)
      • MPSD (multi parameter spectrogram discriminator): like in UnivNet, but we let it judge both phase part and magnitude part
      • PHD (phase discriminator): self-designed, care more about phase, since l_mstft has already regulated magnitude
        • input of MPSD is [(mag_real, phase_real), (mag_fake, phase_fake)], thus distinguishes real/fake stft data
        • input of PHD is [(mag_real, phase_real), (mag_real, phase_fake)], thus ONLY distinguishes real/fake phase
  • losses
    • l_adv (adversarial loss): modified from HiFiGAN, but relativized
    • l_fm (feature map loss): borrowed from MelGAN
    • l_mstft (multi stft loss): modified from Parallel WaveGAN, but we calculate mel_loss rather than linear_loss
    • l_env (envlope loss): borrowed from RefineGAN
    • l_dyn (dynamic loss): self-designed, inspired by l_env
    • l_sm (strip mirror loss): self-designed, but might hurts audio quality :(

Oh my dude, it's really a biggy feng-he monster :(

Tips of ideas to try or failed:

• To divide consonant part against vowel part in time domain, then use two generator to generate them separately
  • I found this will bring breakups in conjunction points, thus audio sounds noisy, yet mstft loss will be more unstable
• To shallow fuse the reference wav with half-decoded mel, rather than the encode-merge-decode UNet architecture
  • I found this will make the generator lazy to learn, even overfit to train set
  • might because waveform is far from audio semantics but near representation, so an encoder is necessary to extract semantical info
• NOTE: the weight of mstft loss should NOT be too overwhelming in front of adversarial loss (like in HiFiGAN)
  • adversarial loss leads to more clear plosives (b/p/g/k/d/t), while mstft loss contributes little to consonants
  • hop_length in stft is much larger that stride in discriminators, thus mstft loss is usually more coarse than adversarial loss in time domain

Acknowledgements

Codes referred to:

• keithito's Tacotron
• jaywalnut310's MelGAN
• jik876's official HiFiGAN
• ming024's FastSpeech2

Ideas plagiarized from:

• Tacotron: Towards End-to-End Speech Synthesis
• FastSpeech: Fast, Robust and Controllable Text to Speech
• FastSpeech 2: Fast and High-Quality End-to-End Text to Speech
• MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis
• Multi-band MelGAN: Faster Waveform Generation for High-Quality Text-to-Speech
• HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis
• UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation
• RefineGAN: Universally Generating Waveform Better than Ground Truth with Highly Accurate Pitch and Intensity Responses

code release kept under the MIT license, greatest thanks all the authors!! :)

---

by Armit 2022/02/15 2022/05/25