Update README.md

README.md

@@ -15,3 +15,77 @@ configs:
   - split: train
     path: data/train-*
 ---
+
+# Dataset Card for MyrtleNoiseData
+
+This dataset provides background noise audio, suitable for noise augmentation
+while training ASR models.
+
+## Dataset Details
+
+### Dataset Description
+
+Curated by: [Myrtle.ai](https://myrtle.ai/)
+
+License: Myrtle.ai's modifications to the source data are licensed under
+  the [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) license.
+  Some of the original data is under the [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/) license; the rest is in the public domain. 
+  Please see the Source Data section below for more information.
+
+## Uses
+
+The noise audio is intended to be combined with speech audio at 
+signal-to-noise ratios in the range 0--60 dB.
+
+## Dataset Structure
+
+This dataset contains 1155 audios, all in the train split.
+
+You can access the first audio like this:
+```python
+>>> import datasets
+>>> noise = datasets.load_dataset("TODO/MyrtleNoiseData")
+>>> noise["train"][0]["audio"]["array"]
+array([-0.17913818, -0.26080322, -0.1835022 , ..., -0.26644897,
+       -0.2434082 , -0.25830078])
+```
+All of the data is 16 kHz and single-channel.
+
+
+## Dataset Creation
+
+### Source Data
+
+- 843 of the audios originate from
+[Free Sound](https://www.freesound.org), 
+as collected for the [MUSAN](https://www.openslr.org/17/) dataset. All these audios are in the public domain.
+- The remaining 312 audios were collected from YouTube videos marked as [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/).
+  Specific attributions are [here](./youtube_attributions.md)
+
+
+#### Data Collection and Processing
+
+Any audio with understandable human speech was filtered out.
+
+Random 20s segments of the YouTube audio were selected.
+
+
+#### Personal and Sensitive Information
+
+Contains no personal information
+
+## Bias, Risks, and Limitations
+
+This dataset contains a large variety of background noises, but not all
+types of background noise are included. If your target validation dataset
+has a type of background noise not included here, then using this 
+noise dataset for augmentation may not help.
+
+If your training dataset already contains significant amounts of 
+background noise, then training with noise augmentation may not be
+necessary.
+
+
+## Dataset Card Contact
+
+hello@myrtle.ai