Official implementation of EdiTTS: Score-based Editing for Controllable Text-to-Speech. Audio samples are available on our demo page.

We present EdiTTS, an off-the-shelf speech editing methodology based on score-based generative modeling for text-to-speech synthesis. EdiTTS allows for targeted, granular editing of audio, both in terms of content and pitch, without the need for any additional training, task-specific optimization, or architectural modifications to the score-based model backbone. Specifically, we apply coarse yet deliberate perturbations in the Gaussian prior space to induce desired behavior from the diffusion model, while applying masks and softening kernels to ensure that iterative edits are applied only to the target region. Listening tests demonstrate that EdiTTS is capable of reliably generating natural-sounding audio that satisfies user-imposed requirements.

Please cite this work as follows.

@misc{tae&kim2021editts,
      title={EdiTTS: Score-based Editing for Controllable Text-to-Speech}, 
      author={Jaesung Tae and Hyeongju Kim and Taesu Kim},
      year={2021}
}

1. Create a Python virtual environment (venv or conda) and install package requirements as specified in requirements.txt.

   python -m venv venv
   source venv/bin/activate
   pip install -U pip
   pip install -r requirements.txt

2. Build the monotonic alignment module.

   cd model/monotonic_align
   python setup.py build_ext --inplace

For more information, refer to the official repository of Grad-TTS.

The following checkpoints are already included as part of this repository, under checkpts.

• Grad-TTS (old ver.)
• HiFi-GAN (LJ_FT_T2_V1 ver.)

1. Prepare an input file containing samples for speech generation. Mark the segment to be edited via a vertical bar separator, |. For instance, a single sample might look like

   In | the face of impediments confessedly discouraging |

   We provide a sample input file in resources/filelists/edit_pitch_example.txt.

2. To run inference, type

   CUDA_VISIBLE_DEVICES=0 python edit_pitch.py \
       -f resources/filelists/edit_pitch_example.txt \
       -c checkpts/grad-tts-old.pt -t 1000 \
       -s out/pitch/wavs

   Adjust CUDA_VISIBLE_DEVICES as appropriate.

1. Prepare an input file containing pairs of sentences. Concatenate each pair with # and mark the parts to be replaced with a vertical bar separator. For instance, a single pair might look like

   Three others subsequently | identified | Oswald from a photograph. #Three others subsequently | recognized | Oswald from a photograph.

   We provide a sample input file in resources/filelists/edit_content_example.txt.

2. To run inference, type

   CUDA_VISIBLE_DEVICES=0 python edit_content.py \
       -f resources/filelists/edit_content_example.txt \
       -c checkpts/grad-tts-old.pt -t 1000 \
       -s out/content/wavs

• Grad-TTS
• HiFi-GAN
• SDEdit

Released under the modified GNU General Public License.