iqbaltaha3/indic-voices-indian-language-classification

Indic Language Speech Classification Dataset

A curated subset of the AI4Bharat IndicVoices-R dataset created for Indian-language speech classification experiments on consumer hardware.

The original IndicVoices-R corpus is extremely large (~750 GB). This dataset intentionally samples a much smaller subset so that researchers and developers can download, preprocess, and experiment with the data locally on machines with limited storage and memory.

Languages

The dataset contains six Indian languages:

• Hindi
• Bengali
• Marathi
• Tamil
• Telugu
• Kannada

Dataset construction

The subset was created using the following pipeline:

1. Select the six target languages.
2. Download IndicVoices-R parquet shards incrementally.
3. Apply quality filtering.
4. Identify speakers with sufficient high-quality recordings.
5. Randomly sample speakers.
6. Randomly sample 15 clips per selected speaker.
7. Split speakers into train/validation/test sets.
8. Save the resulting curated dataset.

The sampling process uses a fixed random seed (42) to make the selection reproducible.

Quality filtering

A recording is retained only when:

Criterion	Threshold
Duration	3–15 seconds
SNR	≥ 15
CER	≤ 0.15
Speaker ID	Required

The purpose of these filters is to remove recordings that are too short, too long, noisy, poorly transcribed, or unsuitable for speaker-level experiments.

Speaker-disjoint splits

The dataset is split by speaker rather than by individual audio clip.

Approximately:

• 70% speakers → training
• 15% speakers → validation
• 15% speakers → test

Therefore, speakers appearing in the training set do not appear in the validation or test sets.

This is important because it allows evaluation of how well the model generalizes to previously unseen speakers.

Dataset size

Language	Train	Validation	Test
Hindi	840	180	180
Bengali	840	180	180
Marathi	840	180	180
Tamil	605	120	150
Telugu	630	135	135
Kannada	840	180	180

The smaller Tamil and Telugu splits are a consequence of the quality filtering and minimum-speaker-clip requirements.

Speaker sampling

Target speakers:

Language	Target
Hindi	80
Bengali	80
Marathi	80
Tamil	80
Telugu	60
Kannada	80

The target is not guaranteed. If fewer eligible speakers are available after quality filtering, all eligible speakers are used.

Each selected speaker contributes 15 clips.

Features

The dataset contains the following metadata:

• language
• lang
• speaker_id
• split
• audio
• duration
• snr
• cer
• gender
• age_group
• scenario
• state

The audio field contains the original WAV bytes.

Audio preprocessing is intentionally not performed in this phase.

Intended use

This dataset is intended primarily for:

• Indian-language speech classification
• Speaker-independent classification experiments
• CNN / ResNet experiments
• Audio representation learning
• Benchmarking lightweight models
• Local experimentation on consumer hardware

It is particularly useful when the full IndicVoices-R corpus is too large for local experimentation.

Limitations

This dataset is a curated subset and should not be considered a representative sample of the complete IndicVoices-R corpus.

The language distribution is not perfectly balanced because the number of eligible speakers depends on the quality filters.

The dataset should therefore be used primarily for experimentation and proof-of-concept research rather than population-level conclusions.

Source

Original dataset:

AI4Bharat IndicVoices-R

https://huggingface.co/datasets/ai4bharat/indicvoices_r

This dataset is derived from the original corpus. Please refer to the original dataset documentation and license for attribution and usage requirements.

Reproducibility

Sampling seed:

42

The data acquisition and curation code is available in the accompanying GitHub repository.

Project pipeline

This dataset represents Phase 1 of the project:

Raw IndicVoices-R ↓ Speaker selection ↓ Quality filtering ↓ Curated dataset ↓ Mel-spectrogram preprocessing ↓ CNN / ResNet / PANNs / Whisper experiments