Equal AI org

Dataset Name: Hindi ASR Merged 8k (Male + Female)

Description
This dataset is a consolidated Hindi Automatic Speech Recognition (ASR) corpus created by merging two publicly available datasets: All Hindi ASR Male v1.1 and All Hindi ASR Female v1.1. The merged dataset follows a randomized integration strategy to ensure balanced speaker distribution and improved diversity for speech recognition model training.

All audio samples have been uniformly downsampled to 8 kHz to support telephony-grade ASR use cases, low-bandwidth speech systems, and edge deployment scenarios where compact audio representation is essential.

The final dataset is approximately 30 GB in size and is designed to support research and development in Hindi speech recognition, speech-to-text modeling, and low-resource ASR experimentation.

Source Datasets
This dataset is derived from the following open datasets:

SayantanJoker/All_Hindi_ASR_Male_v1.1

SayantanJoker/All_Hindi_ASR_Female_v1.1

Proper credit goes to the original creators and contributors of these datasets.

Key Features

Language: Hindi

Audio format: WAV (downsampled to 8 kHz)

Total size: ~30 GB

Speaker coverage: Male and Female speakers

Sampling strategy: Randomized merge across both source datasets

Use cases:

Automatic Speech Recognition (ASR)

Speech-to-text model training

Low-bandwidth and telephony ASR systems

Accent and speaker diversity research

Preprocessing Details

Audio from both source datasets was randomly merged to avoid ordering or speaker bias.

All audio files were downsampled to 8 kHz for consistency and optimized ASR experimentation in constrained environments.

Original transcripts and metadata were preserved wherever available.

File naming and structure were normalized for easier ingestion into ASR pipelines.

Intended Use
This dataset is intended for:

Training and fine-tuning Hindi ASR models

Benchmarking speech recognition systems

Research in multilingual and low-resource ASR

Edge and telephony speech applications

Limitations

Downsampling to 8 kHz may reduce performance for models expecting higher fidelity audio (16 kHz or above).

Quality and transcription accuracy depend on the original source datasets.

Speaker demographics and recording environments are inherited from the original datasets.

pavanmantha changed pull request status to merged

Sign up or log in to comment