README.md

metadata

language:
  - en

Marathi Sentence-Level ASR Dataset

📌 Overview

This dataset contains sentence-level Marathi speech segments aligned with transcripts.

The dataset was created by extracting subtitle timestamps (SRV3 format) from Marathi YouTube content and segmenting the corresponding audio using precise time alignment.

Each sample contains:

• A WAV audio file (sentence-level)
• The corresponding Marathi transcript text

This dataset is suitable for:

• Whisper fine-tuning
• Wav2Vec2 CTC training
• Marathi speech recognition research
• ASR benchmarking

---

📊 Dataset Statistics

• Language: Marathi (mr)
• Alignment Level: Sentence-level
• Format: WAV (mono)
• Metadata format: Hugging Face audiofolder
• Number of samples: ~8,900+
• Segment duration range: ~2–20 seconds

Exact dataset size may vary depending on filtering thresholds.

---

📂 Dataset Structure

HF_Dataset_Sentences/ └── train/ ├── audio/ │ ├── 00001.wav │ ├── 00002.wav │ └── ... └── metadata.csv

metadata.csv format

file_name,text audio/00001.wav,नमस्कार मी भाऊ तोरसेकर audio/00002.wav,आज आपण चर्चा करणार आहोत

🛠 Data Creation Process

1. Audio downloaded from publicly available Marathi YouTube content.
2. Subtitles extracted in SRV3 format.
3. Sentence blocks (<p> tags) parsed for start and duration.
4. Audio segmented using timestamp alignment.
5. Text normalized and saved in Hugging Face audiofolder format.

---

🎯 Intended Use

This dataset is intended for:

• Academic research
• Marathi ASR system development
• Speech model fine-tuning
• Low-resource language modeling