File size: 2,812 Bytes
3b7cb69 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 |
---
license: mit
---
`TrainingSpeech` is an initiative to provide **open and freely reusable dataset** of voices
- for speech-to-text models training
- on non-english languages
- using already available data (such as audio-books).
Right now, data are extracted exclusively from audio-books and in French language. Let me know if you are intersted to contribute [by creating an issue](https://github.com/wasertech/TrainingSpeech/issues/new).
## Tooling
`TrainingSpeech` comes with a CLI that automate and simplify:
- transcript extraction
- [forced-alignment](https://github.com/pettarin/forced-alignment-tools#definition-of-forced-alignment) (using [aeneas](https://github.com/readbeyond/aeneas))
- validation and correction
## Common workflow
### 1. Generate and validate alignment on existing source
1. pick a source that have NOT been validated yet: see `python manage.py stats` and `./sources.json` for more info
2. download assets (ie epub and mp3 files): `python manage.py download -s <SOURCE_NAME>`
3. check alignment: `python manage.py check-alignment <SOURCE_NAME>` (may require multiple iterations)
4. send a pull request with generated transcript and alignment
### 2. Add New source (team members only)
1. retrieve epub and corresponding mp3 file and store them into `./data/epubs` and `./data/mp3` (respectively)
2. create new source into `./sources.json` (NB: all fields are mandatory)
3. generate initial transcript using `python manage.py build-transcript <SOURCE_NAME>`
4. upload epub and mp3 files on S3 `python manage.py upload -s <SOURCE_NAME>`
## Dev setup
```sh
$ sudo apt-get install -y ffmpeg espeak libespeak-dev python3-numpy python-numpy libncurses-dev libncursesw5-dev sox libsqlite3-dev
$ git clone [email protected]:wasertech/TrainingSpeech.git
$ pip3 install --user pipenv
$ cd TrainingSpeech
$ pipenv install --python=3.6.6
$ pipenv sync
$ pipenv shell
$ pytest
```
## Last releases & download
Releases are ready-to-use `zip` archives containing :
- short 16kHz 16bit wav audio speeches (0-15s)
- a single `data.csv` file with following columns:
- `path`: path to the audio file inside the archive
- `duration`: audio duration in second
- `text`: transcript
| Name | # speeches | # speakers | Total Duration | Language |
|:--------------------------------------------------------------------------------------------------------|-------------:|-------------:|:---------------|:-----------|
| [2019-04-11_fr_FR](https://huggingface.co/datasets/wasertech/TrainingSpeech/resolve/main/ts_2019-04-11_fr_FR.zip?download=true) (w/ 💖 from [@lissyx](https://github.com/lissyx)) | 124089 | 4 | 182:43:35 | fr_FR | |