January 18, 2019

633 words 3 mins read



A method to generate speech across multiple speakers

repo name facebookarchive/loop
repo link https://github.com/facebookarchive/loop
language Python
size (curr.) 146 kB
stars (curr.) 836
created 2017-08-16
license Other


PyTorch implementation of the method described in the paper VoiceLoop: Voice Fitting and Synthesis via a Phonological Loop.

VoiceLoop is a neural text-to-speech (TTS) that is able to transform text to speech in voices that are sampled in the wild. Some demo samples can be found here.

Quick Start

Follow the instructions in Setup and then simply execute:

python generate.py  --npz data/vctk/numpy_features_valid/p318_212.npz --spkr 13 --checkpoint models/vctk/bestmodel.pth

Results will be placed in models/vctk/results. It will generate 2 samples:

You can also generate the same text but with a different speaker, specifically:

python generate.py  --npz data/vctk/numpy_features_valid/p318_212.npz --spkr 18 --checkpoint models/vctk/bestmodel.pth

Which will generate the following sample.

Here is the corresponding attention plot:

Legend: X-axis is output time (acoustic samples) Y-axis is input (text/phonemes). Left figure is speaker 10, right is speaker 14.

Finally, free text is also supported:

python generate.py  --text "hello world" --spkr 1 --checkpoint models/vctk/bestmodel.pth


Requirements: Linux/OSX, Python2.7 and PyTorch 0.1.12. Generation requires installing phonemizer, follow the setup instructions there. The current version of the code requires CUDA support for training. Generation can be done on the CPU.

git clone https://github.com/facebookresearch/loop.git
cd loop
pip install -r scripts/requirements.txt


The data used to train the models in the paper can be downloaded via:

bash scripts/download_data.sh

The script downloads and preprocesses a subset of VCTK. This subset contains speakers with american accent.

The dataset was preprocessed using Merlin - from each audio clip we extracted vocoder features using the WORLD vocoder. After downloading, the dataset will be located under subfolder data as follows:

├── data
    └── vctk
        ├── norm_info
        │   ├── norm.dat
        ├── numpy_feautres
        │   ├── p294_001.npz
        │   ├── p294_002.npz
        │   └── ...
        └── numpy_features_valid

The preprocess pipeline can be executed using the following script by Kyle Kastner: https://gist.github.com/kastnerkyle/cc0ac48d34860c5bb3f9112f4d9a0300.

Pretrained Models

Pretrainde models can be downloaded via:

bash scripts/download_models.sh

After downloading, the models will be located under subfolder models as follows:

├── data
├── models
    ├── blizzard
    ├── vctk
    │   ├── args.pth
    │   └── bestmodel.pth
    └── vctk_alt

Update 10/25/2017: Single speaker model available in models/blizzard/


Finally, speech generation requires SPTK3.9 and WORLD vocoder as done in Merlin. To download the executables:

bash scripts/download_tools.sh

Which results the following sub directories:

├── data
├── models
├── tools
    ├── SPTK-3.9
    └── WORLD



Single speaker model is trained on blizzard 2011. Data should be downloaded and prepared as described above. Once the data is ready, run:

python train.py --noise 1 --expName blizzard_init --seq-len 1600 --max-seq-len 1600 --data data/blizzard --nspk 1 --lr 1e-5 --epochs 10

Then, continue training the model with :

python train.py --noise 1 --expName blizzard --seq-len 1600 --max-seq-len 1600 --data data/blizzard --nspk 1 --lr 1e-4 --checkpoint checkpoints/blizzard_init/bestmodel.pth --epochs 90


Training a new model on vctk, first train the model using noise level of 4 and input sequence length of 100:

python train.py --expName vctk --data data/vctk --noise 4 --seq-len 100 --epochs 90

Then, continue training the model using noise level of 2, on full sequences:

python train.py --expName vctk_noise_2 --data data/vctk --checkpoint checkpoints/vctk/bestmodel.pth --noise 2 --seq-len 1000 --epochs 90


If you find this code useful in your research then please cite:

  title           = {VoiceLoop: Voice Fitting and Synthesis via a Phonological Loop},
  author          = {Taigman, Yaniv and Wolf, Lior and Polyak, Adam and Nachmani, Eliya},
  journal         = {ArXiv e-prints},
  archivePrefix   = "arXiv",
  eprinttype      = {arxiv},
  eprint          = {1705.03122},
  primaryClass    = "cs.CL",
  year            = {2017}
  month           = October,


Loop has a CC-BY-NC license.

comments powered by Disqus