EmoSpeech: Guiding FastSpeech2 Towards Emotional Text to Speech

How to run

Build env

You can build an environment with Docker or Conda.

To set up environment with Docker

If you don't have Docker installed, please follow the links to find installation instructions for Ubuntu, Mac or Windows.

Build docker image:

docker build -t emospeech .

Run docker image:

bash run_docker.sh

To set up environment with Conda

If you don't have Conda installed, please find the installation instructions for your OS here.

  conda create -n etts python=3.10
  conda activate etts
  pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
  pip install -r requirements.txt

If you have different version of cuda on your machine you can find applicable link for pytorch installation here.

Download and preprocess data

We used data of 10 English Speakers from ESD dataset. To download all .wav, .txt files along with .TextGrid files created using MFA:

  bash download_data.sh

To train a model we need precomputed durations, energy, pitch and eGeMap features. From src directory run:

  python -m src.preprocess.preprocess

This is how your data folder should look like:

  .
  ├── data
  │   ├── ssw_esd
  │   ├── test_ids.txt
  │   ├── val_ids.txt
  └── └── preprocessed
          ├── duration
          ├── egemap
          ├── energy
          ├── mel
          ├── phones.json
          ├── pitch
          ├── stats.json
          ├── test.txt
          ├── train.txt
          ├── trimmed_wav
          └── val.txt

Training

1. Configure arguments in config/config.py.
2. Run python -m src.scripts.train.

Testing

Testing is implemented on testing subset of ESD dataset. To synthesize audio and compute neural MOS (NISQA TTS): 1. Configure arguments in config/config.py under Inference section. 2. Run python -m src.scripts.test.

You can find NISQA TTS for original, reconstructed and generated audio in test.log.

Inference

EmoSpeech is trained on phoneme sequences. Supported phones can be found in data/preprocessed/phones.json. This repositroy is created for academic research and doesn't support automatic grapheme-to-phoneme conversion. However, if you would like to synthesize arbitrary sentence with emotion conditioning you can: 1. Generate phoneme sequence from graphemes with MFA.

  1.1 Follow the [installation guide](https://montreal-forced-aligner.readthedocs.io/en/latest/installation.html)

  1.2 Download english g2p model: `mfa model download g2p english_us_arpa`

  1.3 Generate phoneme.txt from graphemes.txt: `mfa g2p graphemes.txt english_us_arpa phoneme.txt`

1. Run python -m src.scripts.inference, specifying arguments:

Аrgument | Meaning | Possible Values | Default value ---|---|---|--- -sq | Phoneme sequence to synthesisze| Find in data/phones.json. | Not set, required argument. -emo | Id of desired voice emotion | 0: neutral, 1: angry, 2: happy, 3: sad, 4: surprise. | 1 -sp | Id of speaker voice | From 1 to 10, correspond to 0011 ... 0020 in original ESD notation. | 5 -p | Path where to save synthesised audio | Any with .wav extension. | generationfromphoneme_sequence.wav

For example

python -m src.scripts.inference --sq "S P IY2 K ER1 F AY1 V T AO1 K IH0 NG W IH0 TH AE1 NG G R IY0 IH0 M OW0 SH AH0 N"

If result file is not synthesied, check inference.log for OOV phones.

References

1. FastSpeech 2 - PyTorch Implementation
2. iSTFTNet : Fast and Lightweight Mel-spectrogram Vocoder Incorporating Inverse Short-time Fourier Transform
3. Publicly Available Emotional Speech Dataset (ESD) for Speech Synthesis and Voice Conversion
4. NISQA: Speech Quality and Naturalness Assessment
5. Montreal Forced Aligner Models
6. Modified VocGAN
7. AdaSpeech