|
--- |
|
license: cc0-1.0 |
|
language: |
|
- en |
|
task_categories: |
|
- text-to-speech |
|
- text-generation |
|
pretty_name: Common Voice 11 (en) Cleaned and Tokenized |
|
size_categories: |
|
- 10K<n<100K |
|
--- |
|
|
|
A cleaned and tokenized version of the English data from [Mozilla Common Voice 11 dataset](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/tree/main). |
|
|
|
Cleaning steps: |
|
* Filtered on samples with >2 upvotes and <1 downvotes] |
|
* Removed non voice audio at start and end through pytorch VAD |
|
|
|
Tokenization |
|
* Audio tokenized through [EnCodec by Meta](https://github.com/facebookresearch/encodec) |
|
* Using 24khz pre-trained model, and target bandwidth of 1.5 |
|
* Represented in text as audio_token_0 - audio_token_1023 |
|
* Prompts constructed as "text: \<common voice transcript\>\naudio: \<audio tokens\>" |
|
* Prompts tokenized with GPT tokenizer with added vocab of audio tokens. |
|
* Tokenized prompts padded to size 1024 with eos_token. |
|
|
|
Each sample has 3 properties: input_ids, attention_mask and labels. input_ids and labels are the tokenized prompts and attention_mask is the attention mask. |
