A Speech-to-Text dataset scraped from NPTEL for Indo-English accent, from Education Domain.

The dataset was not manually annotated by us. We assume NPTEL has used Google ASR on top of which they have made reasonable amount of corrections.

We split the dataset as follows: (randomly sampled)

Note:

• Sample Set is a small subset manually annotated by us to compute the quality of data. We refer to it as Pure Set.
• We computed (in May 2020) the following results using the Pure Set for benchmarking purposes.
  • All the DL models listed below uses LM.
  • For Jasper and QuartzNet, we use models from Nvidia's NeMo framework.
  • For DeepSpeech, we use Baidu's official model.
  • The benchmarking code for others can be found here.

To understand if the data we have crawled is useful, we sample 500k chunks from the train set and fine-tune an English ASR model using that for an epoch. We chose QuartzNet (Ultra-Tiny) pre-trained model for the fine-tuning because it was lightweight as well as very competitive in accuracy as seen above. (Training code can be found in this same repo)

On the Pure-Set, we observed an improvement in WER from 0.5034 (pre-trained model without LM) to 0.3207 (fine-tuned model without LM), which signified a promising scope to use the crawled dataset for much further improvement.

The pre-mature fine-tuned model can be found in the GitHub Releases section of the repo. (Currently due to lack of compute, we couldn't exhaustively use the data and find out results across different models & methods)

• Even though the dataset is noisy compared to publicly available datasets, we believe it would serve as a good intial data for building models.
• Especially this dataset focuses on South Asian English accent, and is of education domain.
• Even the raw audio from this dataset would be useful for pre-training ASR models like Wav2Vec 2.0.
  (As can be seen on this recent leaderboard)
• For a better but closed dataset, check this recent competition: IIT-M Speech Lab - Indian English ASR Challenge
  (could be available on request)

• Sample Data (Pure-Set)
• Train, Test and Dev sets downloader scripts

In some cases, we might need data containing a single speaker(TTS, Speaker recognition, etc). For that, choose a youtube playlist of your choice and crawl it.

Please check here for the instructions to do that on the crawler we used.

For clarifications, please write on the GitHub Issues section.