A tool that makes creating text-to-speech corpora fairly easy. It was developed and tested under Windows, I heard that the GUI backend might be a little more difficult to install on other systems. There is an example setup that I used myself called ExampleCorpus. It uses a subset of the prompts of the preview of the second iteration of the Thorsten corpus.
- Create an environment (here shown with conda as example)
conda create --prefix ./toucan_conda_venv --no-default-packages python=3.8
- Install dependencies:
pip install --no-cache-dir -r requirements.txt
To start a new corpus, create a directory inside the Corpora directory and give it the name you desire. In this
directory, create a file called prompts.txt, which should contain one utterance per line. Update
the current_corpus_name property of the parameters.py file. You should update the other settings as well, while you
are there.
To record a corpus, run CorpusCreator.py. If you started on it in a prior session, it will load and continue where you
left off. If you start it for the first time, it will create the file that maps prompts to audios and begin with the
first prompt. A window will open up, which is controlled by keybinds only.
The window will display a sentence. When you are ready, press and hold [CTRL]. While that key is pressed, your microphone will record. After you release the key, the recording is done, the audio will be saved, and the lookup will be updated. Then the next datapoint will be loaded. Then you can record again with [CTRL]. If you are not happy with a recording, you can press [ALT] to go back to the last prompt and overwrite the last one you recorded. You can go back multiple steps, but you have to re-record all the ones you are going back over. When you are done with a session, press [ESC] to exit the program safely.
The format of the corpus is close to LJSpeech style, however it assumes the texts are normalized already.
The recordings are done in 48kHz by default, since they can always be downsampled afterwards, and that should be enough for TTS. For prosodically standard speech, even 16kHz is enough in my opinion. You can do super-resolution with a vocoder.
A bit of signal processing is applied to get a uniform loudness and silences are automatically removed from beginning
and end. Depending on your microphone, you might want to tweak some settings in the code. Just record a few samples and
see if they are fine before committing to something big. Unprocessed audio is also saved in the unprocessed subfolder that
will be created in the corpus directory when you run it first.