MahaDhwani

MahaDhwani is a corpus comprising 279K hours of raw audio across 22 Indian languages and English. We propose a framework to create large raw audio datasets for under-represented languages by collating publicly accessible audio content.

Download

Our Dataflow Pipeline works as follows-

1. Video_ids are fetched from a postgreSQL table which contains all the video_ids, metadata and related info(which is blank intially).
2. Video_ids are assigned to each VM through dataflow.
3. Audios are downloaded on each VM using yt-dlp and then uploaded to the cloud bucket.
4. After successful upload, the PostgreSQL table is updated with the metadata, bucket_path, duration, file size, etc.

Steps to run the pipeline -

1. Setup a PostgreSqL table, a cloud bucket and update the code in pipeline.py accordingly.
2. Setup a GCP account for dataflow access.
3. Create and push the Dockerfile provided for setting up VM environments.
   • Make sure that the apache beam version in the dockerfile matches with the local environment(Python 3.10 and Apache beam sdk 2.58.1 were used here).
4. Run the bash script -

bash run.sh

5. For filtering video_ids based on metadata, dataflow_pipeline/languages/*/video_ids_*.csv files can be used.

Model

Model	Link
MahaDhwani Pretrained Conformer Encoder	HF

Analysis

1. Effect of Pretraining -

We compare the performance of ASR models fine-tuned on the IndicVoices dataset, starting with (i) random initialisation (ii) pretrained checkpoint of an English ASR model, Nvidia-En-SSL (iii) the IndicConformer-HMS model pretrained on MahaDhwani. Each of the above is trained 12.5%, 50% and 100% of the labeled training data.

2. Comparisons with other models -

We compare IndicConformer-HMS(pretrained with MahaDhwani) with existing massively multilingual models, namely, USM, Whisper and MMS. We find that it significantly outperforms all existing models.