Wazobia enables machines to understand African english better.
# [![wazobia logo][]][play link]
wazobia-open-speech is an automatic speech recognition system that can understand english accents spoken by native Africans. Wazobia open speech is built on top of Mozilla's [DeepSpeech engine][deepspeech link].
## Steps to train
Download [training audio files][wazobia folder link]. Use download as zip option.
Move unzipped wazobia directory to ./util
Change directory to root of repo
Install all python and runtime dependencies
```bash
$ sudo bash prepare.sh
```
Change directory to deepspeech
```bash
$ cd DeepSpeech9
```
Train
```bash
$ python DeepSpeech.py --n_hidden 2048 --checkpoint_dir fine_tuning_checkpoints/ --epochs 100 --train_files training_csvs/train.csv --dev_files training_csvs/dev.csv --test_files training_csvs/test.csv --learning_rate 0.0001 --export_dir output_models/ --use_allow_growth true --train_cudnn true
```
## Tips before training
- Wazobia open speech wiil only run on a linux based system. It has been tested on Ubuntu 18.04.
- Using a GPU is 20x faster, to opt-out of using a GPU set the --train_cudnn flag to false
- We recommend using google colab especially if you don't have a GPU, see ./wzb_colab_example.ipynb
## Benchmark
We provide a [model][benchmark model link] as benchmark trained with the following flags:
- **n_hidden** (number of hidden layers): **2048**
- **epochs**: **100**
- **learning_rate**: **0.0001**
### Result in benchmark
WER = Word Error Rate
CER = Character Error Rate
| WER | CER |
| ------ | ------ |
| 0.992230 | 0.581740 |
## Data collection
Speech data is collected by means of a mobile application ([wazobia open speech mobile][play link]) written with flutter. The app is currently only available for download in Nigeria although a website will be available in the near future for data collection. The source codes for the mobile and web applications will also be availed to the public.
## Collecting your own data
Wazobia open speech mobile and web can be modifie …