sulabhkatiyar/ne-asr-mjw-aug

Version 2.0 (rebuilt 2026-05-29) -- see VERSION HISTORY below

WARNING: This is v2.0, a clean rebuild that supersedes the deprecated v1.0 revision. v1.0 had been built at an unintended 5x expansion (with pitch-shift augmentation on a tonal language) and is retracted. Use v2.0 only.

Version history

v2.0 (HEAD = d721db548b541242ceb6fcb072dbf20618000bca, 2026-05-29) -- current

• Rebuilt at correct 3x augmentation: speed-perturb only (factors 0.9, 1.0, 1.1).
• No pitch shift on this tonal language (pitch shift alters lexical tone contrasts in Karbi; disabled per configs/augmentation_config.yaml and enforced by a belt-and-suspenders override inside scripts/augment_data.py).
• Source: sulabhkatiyar/ne-asr-mjw train split = 508 rows -> 508 x 3 = 1,524 augmented train rows.
• Validation split: 60 rows (copied unchanged from source).
• Test split: 39 rows (copied unchanged from source).
• License: CC-BY-NC 4.0 (inherited from the MMS-1B base model used downstream; the original Vaani recordings are CC-BY-4.0; once combined with MMS-1B-derived training artifacts, downstream use inherits the more restrictive non-commercial term).

v1.0 (PRE-HEAD = e6e30afeb9095af8ad8c84b80d5beb35a3303caf boundary, retracted) -- DO NOT USE

• v1.0 had been built at an unintended 5x expansion (2,540 train rows = 508 x 5), including pitch-shift variants (+/-1 semitone) on top of the speed-perturb variants. This violated the augmentation policy for tonal languages: Karbi has 3 contrastive tones (L/M/H per the Konnerth grammar; tonal_audit.md, 2026-05-29). Pitch shift corrupts tonal contrasts and is therefore disabled for mjw.
• The earlier "mjw non-tonal" classification (e.g. in early Module 6 planning docs) was incorrect; tonal_audit.md is authoritative and lists mjw as tonal.
• Users must NOT use v1.0 (the v1.0 commits have been replaced; this repo was rebuilt fresh via scripts/augment_data.py --clean, so v1.0 parquet files are no longer reachable). This note is preserved for auditability.

License (restate)

This augmented dataset is published under CC-BY-NC 4.0, inherited from the MMS-1B base model used in the downstream fine-tuning pipeline. The original Vaani recordings are CC-BY-4.0; the combined artifact (audio + MMS-derived model outputs) inherits the more restrictive non-commercial term.

---

NE ASR Augmented Dataset -- Karbi (mjw)

Augmented automatic speech recognition dataset for Karbi (mjw), a Tibeto-Burman language spoken in Assam, India.

Source

Augmented from sulabhkatiyar/ne-asr-mjw (original transcribed speech data from the ARTPARK-IISc Vaani project).

Language Information

Property	Value
Language	Karbi
ISO 639-3	mjw
Family	Tibeto-Burman
Region	Assam, India
Tonal	Yes
Tier	B (0.63h original data)

Dataset Statistics

• Original training samples: 508
• Augmented training samples: 1,524 (3x augmentation)
• Train shards: 2
• Estimated original duration: ~0.6 hours
• Estimated augmented duration: ~1.9 hours

Split	Samples
train	1,524
validation	60
test	39

Transformations Applied

Each original training sample produces 3 samples (1 original + 2 speed + 0 pitch):

• Speed perturbation: 0.9x, 1.1x (2 variants per sample)
• Pitch shift: Disabled (tonal language -- pitch shift would alter lexical meaning)
• Noise augmentation: Not applied

SpecAugment Parameters (for training, NOT in this dataset)

These parameters are consumed by the training script and are not baked into the audio files:

• mask_time_prob: 0.05
• mask_time_length: 10
• mask_feature_prob: 0.03
• mask_feature_length: 10
• layerdrop: 0.05

Full augmentation config: configs/augmentation_config.yaml

Dataset Format

• Audio: 16kHz mono WAV (stored as Parquet with audio bytes)
• Text: Transcriptions
• Features: audio, text, language, augmentation
• Augmentation labels: original, speed_0.9, speed_1.1

How to Use

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("sulabhkatiyar/ne-asr-mjw-aug")

# Load only the training split
train = load_dataset("sulabhkatiyar/ne-asr-mjw-aug", split="train")

# Filter to only original (non-augmented) samples
original_only = train.filter(lambda x: x["augmentation"] == "original")

# Filter to a specific augmentation type
speed_09 = train.filter(lambda x: x["augmentation"] == "speed_0.9")

Original Data

• Source dataset: sulabhkatiyar/ne-asr-mjw
• Project: ARTPARK-IISc Vaani
• License: CC-BY-4.0

Citation

If you use this dataset, please cite the Vaani project and acknowledge the augmentation pipeline.