Datasets:

ArchCoder
/

hindi-whisper-chunks

input_features listlengths 80 80	labels listlengths 7 294	length int64 7 294
[[0.6301279067993164,-0.010689020156860352,-0.7618265151977539,-0.829509973526001,-0.457120776176452(...TRUNCATED)	[50258,50276,50359,50363,3941,227,3941,105,31970,17937,3941,104,31881,8485,227,3941,248,27099,3941,2(...TRUNCATED)	230
[[0.4253377318382263,-0.06632959842681885,-0.565270185470581,-0.5116407871246338,-0.2735599279403686(...TRUNCATED)	[50258,50276,50359,50363,3941,227,35082,8703,223,3941,255,3941,113,31970,25411,41858,21981,31970,870(...TRUNCATED)	191
[[0.11975955963134766,-0.2641195058822632,-0.2678724527359009,-0.29506850242614746,-0.34440791606903(...TRUNCATED)	[50258,50276,50359,50363,3941,250,31945,3941,245,46758,31970,17937,49316,3941,104,25411,37139,33926,(...TRUNCATED)	142
[[0.5870498418807983,0.03747981786727905,-0.15087342262268066,-0.23026728630065918,-0.12326860427856(...TRUNCATED)	[50258,50276,50359,50363,3941,103,44500,46758,31881,8485,105,17937,25411,31881,8485,98,17937,31970,2(...TRUNCATED)	120
[[0.5604602098464966,-0.10255444049835205,-0.5848401784896851,-0.24466121196746826,-0.18513572216033(...TRUNCATED)	[50258,50276,50359,50363,44500,17937,31945,8485,97,33926,8485,104,33279,25411,8485,113,44500,17937,3(...TRUNCATED)	165
[[0.4084094166755676,-0.097511887550354,-0.37030649185180664,-0.4693584442138672,-0.3675327301025390(...TRUNCATED)	[50258,50276,50359,50363,3941,227,3941,250,31881,3941,105,49316,17937,8485,228,3941,113,17937,3941,2(...TRUNCATED)	150
[[0.20107388496398926,-0.33648157119750977,-0.520513653755188,-0.5058079957962036,-0.252849578857421(...TRUNCATED)	[50258,50276,50359,50363,3941,94,25411,8485,110,3941,245,36158,17937,37139,43372,8485,110,21981,4185(...TRUNCATED)	219
[[0.5464447736740112,-0.05896484851837158,-0.30980849266052246,-0.24318110942840576,-0.3864728212356(...TRUNCATED)	[50258,50276,50359,50363,3941,238,45938,17937,37139,31881,8485,105,17937,36158,8485,98,17937,8485,97(...TRUNCATED)	201
[[0.6450341939926147,0.033526837825775146,-0.30048930644989014,-0.7758400440216064,-0.47578918933868(...TRUNCATED)	[50258,50276,50359,50363,46758,3941,245,36158,17937,37139,43372,31970,33279,49316,25411,49316,8703,2(...TRUNCATED)	174
[[0.4529878497123718,-0.08090591430664062,-0.32210099697113037,-0.49576497077941895,-0.4059863090515(...TRUNCATED)	[50258,50276,50359,50363,44500,48521,27099,48521,37139,48521,8485,110,33926,3941,245,31970,31881,848(...TRUNCATED)	197

End of preview. Expand in Data Studio

Hindi Whisper Chunks

Preprocessed, feature-extracted audio chunks and labels used to fine-tune ArchCoder/whisper-small-hindi-lora, a LoRA adaptation of Whisper-small for Hindi speech recognition.

Dataset Summary

Raw Hindi audio recordings (approximately 12 minutes each) were segmented into short, Whisper-compatible chunks and converted into model-ready features. This dataset is the output of that preprocessing pipeline: Whisper-format log-mel filterbank features paired with tokenized transcription labels, ready to feed directly into a Whisper training loop without repeating audio processing.

Rows: 5,915
Splits: train (5,122), validation (793)
Size: approximately 1.18 GB
Format: Parquet

Dataset Structure

Each row contains:

Field	Type	Description
`input_features`	list of floats (80-dim)	Log-mel filterbank features extracted from an audio chunk, in Whisper's expected input format
`labels`	list of integers	Tokenized transcription, including Whisper's special tokens (language, task, timestamp markers)
`length`	integer	Token length of the label sequence, used for length-based batching during training

How This Dataset Was Built

Sequence segmentation. Long raw audio files were split into individual spoken segments using timestamped transcript data, since Whisper expects inputs under 30 seconds.
Duration filtering. Chunks longer than 30 seconds or with fewer than 2 characters of transcribed text were dropped, to avoid out-of-memory errors during training and to remove noisy, near-empty labels.
Iterative feature extraction. Mel filterbank features were computed one audio file at a time rather than loading the full raw audio set into memory, since the source recordings were large.
Cached to the Hugging Face Hub. Once processed, the dataset was pushed here so downstream training runs do not need to repeat audio preprocessing.

Intended Use

This dataset is intended for fine-tuning or evaluating Whisper-family speech recognition models on Hindi audio, particularly in low-resource or limited-compute settings. The length field supports length-grouped batching, which reduces wasted padding computation when audio durations vary significantly across a dataset, as they do here.

Limitations

Derived from a single source of raw Hindi audio recordings; may not represent the acoustic diversity (accents, background noise, recording equipment) of Hindi speech more broadly.
The relationship between this dataset and the whisper-small-hindi-lora model, while very likely direct given matching sample counts (5,122 training chunks in both), should be treated as strongly implied rather than independently confirmed from a training config file.
No held-out test split is included here; evaluation of models trained on this data was performed separately against the FLEURS Hindi benchmark.

Model trained on this dataset: ArchCoder/whisper-small-hindi-lora

Downloads last month: 184