Skip to content

Commit

Permalink
Augmentation Documentation (#2355)
Browse files Browse the repository at this point in the history
Augmentation Documentation
  • Loading branch information
cahuja1992 authored and reuben committed Sep 23, 2019
1 parent 7995e42 commit 58fdd55
Showing 1 changed file with 30 additions and 0 deletions.
30 changes: 30 additions & 0 deletions README.md
Expand Up @@ -71,6 +71,7 @@ See the output of `deepspeech -h` for more information on the use of `deepspeech
- [Exporting a model for TFLite](#exporting-a-model-for-tflite)
- [Making a mmap-able model for inference](#making-a-mmap-able-model-for-inference)
- [Continuing training from a release model](#continuing-training-from-a-release-model)
- [Training with Augmentation](#training-with-augmentation)
- [Contribution guidelines](#contribution-guidelines)
- [Contact/Getting Help](#contactgetting-help)

Expand Down Expand Up @@ -418,6 +419,34 @@ python3 DeepSpeech.py --n_hidden 2048 --checkpoint_dir path/to/checkpoint/folder

Note: the released models were trained with `--n_hidden 2048`, so you need to use that same value when initializing from the release models.

### Training with augmentation

Augmentation is a useful technique for better generalization of machine learning models. Thus, a pre-processing pipeline with various augmentation techniques on raw pcm and spectrogram has been implemented and can be used while training the model. Following are the available augmentation techniques that can be enabled at training time by using the corresponding flags in the command line.

#### Audio Augmentation
1. **Standard deviation for Gaussian additive noise:** ```--data_aug_features_additive```
2. **Standard deviation for Normal distribution around 1 for multiplicative noise:** ```--data_aug_features_multiplicative```
3. **Standard deviation for speeding-up tempo. If Standard deviation is 0, this augmentation is not performed:** ```--augmentation_speed_up_std```

#### Spectrogram Augmentation
Inspired by Google Paper on [SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition]( https://arxiv.org/abs/1904.08779)
1. **Keep rate of dropout augmentation on a spectrogram (if 1, no dropout will be performed on the spectrogram)**:
* Keep Rate : ```--augmentation_spec_dropout_keeprate value between range [0 - 1]```

2. **Whether to use frequency and time masking augmentation:**
* Enable / Disable : ```--augmentation_freq_and_time_masking / --noaugmentation_freq_and_time_masking```
* Max range of masks in the frequency domain when performing freqtime-mask augmentation: ```--augmentation_freq_and_time_masking_freq_mask_range eg: 5```
* Number of masks in the frequency domain when performing freqtime-mask augmentation: ```--augmentation_freq_and_time_masking_number_freq_masks eg: 3```
* Max range of masks in the time domain when performing freqtime-mask augmentation: ```--augmentation_freq_and_time_masking_time_mask_rangee eg: 2```
* Number of masks in the time domain when performing freqtime-mask augmentation: ```augmentation_freq_and_time_masking_number_time_masks eg: 3 ```

3. **Whether to use spectrogram speed and tempo scaling:**
* Enable / Disable : ```--augmentation_pitch_and_tempo_scaling / --noaugmentation_pitch_and_tempo_scaling.```
* Min value of pitch scaling: ```--augmentation_pitch_and_tempo_scaling_min_pitch eg:0.95 ```
* Max value of pitch scaling: ```--augmentation_pitch_and_tempo_scaling_max_pitch eg:1.2```
* Max value of tempo scaling: ```--augmentation_pitch_and_tempo_scaling_max_tempo eg:1.2```


## Contribution guidelines

This repository is governed by Mozilla's code of conduct and etiquette guidelines. For more details, please read the [Mozilla Community Participation Guidelines](https://www.mozilla.org/about/governance/policies/participation/).
Expand Down Expand Up @@ -481,3 +510,4 @@ There are several ways to contact us or to get help:
3. [**IRC**](https://wiki.mozilla.org/IRC) - If your question is not addressed by either the [FAQ](https://github.com/mozilla/DeepSpeech/wiki#frequently-asked-questions) or [Discourse Forums](https://discourse.mozilla.org/c/deep-speech), you can contact us on the `#machinelearning` channel on [Mozilla IRC](https://wiki.mozilla.org/IRC); people there can try to answer/help

4. [**Issues**](https://github.com/mozilla/deepspeech/issues) - Finally, if all else fails, you can open an issue in our repo.

0 comments on commit 58fdd55

Please sign in to comment.