83 lines
3.4 KiB
Markdown
83 lines
3.4 KiB
Markdown
# GLaDOS Text-to-speech (TTS) Voice Generator
|
|
|
|
[](https://jenkins.sudo.is/job/ben/job/glados-tts/)
|
|
[](https://git.sudo.is/ben/glados-tts)
|
|
[](https://github.com/benediktkr/glados-tts)
|
|
[](LICENSE)
|
|
|
|
Neural network based TTS Engine.
|
|
|
|
## Notes about this fork
|
|
|
|
Forked by [`ben`](https://git.sudo.is/ben) (:github: [`@benediktkr`](https://github.com/benediktkr)) from
|
|
[`github:VRCWizard/glados-tts-voice-wizard`](https://github.com/VRCWizard/glados-tts-voice-wizard),
|
|
which in turn was a fork of
|
|
[`github:R2D2FISH/glados-tts`](https://github.com/R2D2FISH/glados-tts).
|
|
|
|
This fork modernizes and improves the Python code in the project and does a bunch of housekeeping.
|
|
|
|
* `[DONE]`: Gets rid of the `SciPy` dependency (replaced with the more modern and lightwight [`pysoundfile`](https://github.com/gooofy/py-espeak-ng) (since all it was used for was writing a `.wav` file to disk)
|
|
* `[DONE]`: Support modern stable Python 3 versions, and update dependencies.
|
|
* `[DONE]`: Versioned packages with `poetry` and `pyproject.toml`
|
|
* `[DONE]`: Configuration handling with `click`.
|
|
* `[DONE]`: Better logging with `loguru`
|
|
* `[DONE]`: Python coding style and code quality improvements (proper handling of `file` object, improved logging..)
|
|
* `[DONE]`: Switch to using ASGI with `uvicorn` and `fastapi` instead of Flask and WSGI, and support production-capable deployments as default.
|
|
* `[DONE]`: Docker support
|
|
* `[TODO]`: Support Home Assistant through the [`notify` integration](https://www.home-assistant.io/integrations/notify/)
|
|
* `[TODO]`: see if its possible to avoid `espeak-ng` as a system package dependency (python bindings, buliding the C library, etc)
|
|
|
|
No work on the speech model itself is expected.
|
|
|
|
|
|
|
|
## Description
|
|
The initial, regular Tacotron model was trained first on LJSpeech, and
|
|
then on a heavily modified version of the [Ellen
|
|
McClain](https://en.wikipedia.org/wiki/Ellen_McLain) dataset (all
|
|
non-Portal 2 voice lines removed, punctuation added).
|
|
|
|
* The Forward Tacotron model was only trained on about 600 voice lines.
|
|
* The HiFiGAN model was generated through transfer learning from the sample.
|
|
* All models have been optimized and quantized.
|
|
|
|
## Install
|
|
|
|
First you need to [install the `espeak-ng` system
|
|
packages](https://github.com/espeak-ng/espeak-ng/blob/master/docs/guide.md).
|
|
|
|
```shell
|
|
# for debian/ubuntu:
|
|
sudo apt-get install espeak-ng
|
|
|
|
# for fedora/amazon:
|
|
sudo yum install espeak-ng
|
|
```
|
|
|
|
This can hopefully be improved in the future. There is a Python
|
|
bindings for `espeak` (at a glance, found
|
|
[`py-espeak-ng`](https://github.com/gooofy/py-espeak-ng)).
|
|
|
|
Then install the poetry-managed virtualenv
|
|
|
|
```shell
|
|
poetry install
|
|
```
|
|
|
|
|
|
## Usage
|
|
|
|
If you want to just play around with the TTS, works on the shell:
|
|
|
|
```shell
|
|
poetry run gladosctl
|
|
```
|
|
|
|
The TTS engine can also run as a web server:
|
|
|
|
```shell
|
|
poetry run gladosctl restapi
|
|
```
|
|
|
|
A public instance of the http api is running at `http://www.sudo.is/api/glados`, where you can also read [api documentation](https://www.sudo.is/api/glados/docs).
|