dhruvkys/11k-asr

Hindi ASR 11k

A Hindi automatic speech recognition (ASR) dataset of ~10,000 short read-speech utterances with ground-truth Devanagari transcriptions, derived from Mozilla Common Voice (Hindi). It is a convenient, dependency-light corpus for fine-tuning and WER benchmarking of Whisper and other multilingual / Indic ASR models — large enough to move the needle when fine-tuning, small enough to iterate on a single GPU.

• Language: Hindi (hi), Devanagari script
• Task: Automatic Speech Recognition (speech → text)
• Audio: MP3, mono
• Splits: train (9,000 clips) / test (1,000 clips)
• Source: Mozilla Common Voice Hindi (crowd-sourced read speech)
• License: CC0-1.0 (public domain), inherited from Common Voice

Dataset structure

Each split contains an audio folder and a metadata file that maps every clip to its transcription. Every metadata row has a matching audio file.

11k-asr/
├── train/
│   ├── clips/         # 9,000 common_voice_hi_*.mp3
│   └── metadata.tsv   # file_name, text
└── test/
    ├── clips/         # 1,000 common_voice_hi_*.mp3
    └── metadata.tsv   # file_name, text

Data fields

• file_name (string): audio filename, relative to the split's clips/ folder.
• text (string): the ground-truth Hindi transcription in Devanagari.
• audio (Audio): decoded waveform + sampling rate, once loaded (see below).

Data instance

{
  "file_name": "common_voice_hi_27545949.mp3",
  "text": "जिम मेरे घर में ठहरा था।"
}

Splits

Split	Clips
train	9,000
test	1,000
total	10,000

Usage

The metadata is tab-separated and its file_name column is relative to the clips/ folder, so the most reliable way to load audio + text together is to download the repo and join the paths:

import os
import pandas as pd
from huggingface_hub import snapshot_download
from datasets import Dataset, Audio

root = snapshot_download("dhruvkys/11k-asr", repo_type="dataset")

def load_split(split):
    df = pd.read_csv(os.path.join(root, split, "metadata.tsv"), sep="\t")
    df["audio"] = df["file_name"].apply(
        lambda f: os.path.join(root, split, "clips", f)
    )
    return Dataset.from_pandas(df).cast_column("audio", Audio(sampling_rate=16_000))

train = load_split("train")
test  = load_split("test")

print(train[0]["text"])
print(train[0]["audio"]["array"].shape, train[0]["audio"]["sampling_rate"])

Just need the transcripts? Read the metadata directly:

import pandas as pd

df = pd.read_csv(
    "hf://datasets/dhruvkys/11k-asr/train/metadata.tsv",
    sep="\t",                     # metadata is tab-separated
)
print(df.head())

Evaluating a Whisper model (WER)

import evaluate
from transformers import pipeline

asr = pipeline("automatic-speech-recognition",
               model="openai/whisper-small", generate_kwargs={"language": "hi"})
wer = evaluate.load("wer")

preds = [asr(x["audio"])["text"] for x in test]
refs  = [x["text"] for x in test]
print("WER:", wer.compute(predictions=preds, references=refs))

Source and collection

Audio and transcriptions originate from the Mozilla Common Voice Hindi corpus — sentences read aloud and contributed by volunteers, then validated by community review. This dataset is a curated subset repackaged for convenient fine-tuning and evaluation; no new recordings were made.

Preprocessing recommendations

• Common Voice clips are typically 8–48 kHz MP3; resample to 16 kHz mono for Whisper and most ASR models (shown above via Audio(sampling_rate=16000)).
• Transcriptions preserve original casing, punctuation, and occasional Latin-script tokens (e.g. brand/product names in headlines). For WER, consider normalizing punctuation and Latin numerals/tokens depending on your evaluation protocol.

Licensing

Released under CC0-1.0 (public domain dedication), consistent with the Mozilla Common Voice license. You may use, modify, and redistribute the data, including commercially, without restriction. Attribution to Mozilla Common Voice is appreciated but not required.

Citation

If you use this dataset, please cite Mozilla Common Voice:

@inproceedings{commonvoice,
  author    = {Ardila, R. and Branson, M. and Davis, K. and Henretty, M. and
               Kohler, M. and Meyer, J. and Morais, R. and Saunders, L. and
               Tyers, F. M. and Weber, G.},
  title     = {Common Voice: A Massively-Multilingual Speech Corpus},
  booktitle = {Proceedings of the 12th Conference on Language Resources and
               Evaluation (LREC 2020)},
  pages     = {4211--4215},
  year      = {2020}
}

Known limitations

• Modest scale. ~10k clips — strong for fine-tuning and benchmarking, but not a substitute for the full Common Voice corpus when training from scratch.
• Domain skew. Many sentences are news headlines, so vocabulary leans toward named entities, politics, and current-affairs terms.
• Metadata format. Metadata files are tab-separated (.tsv); pass sep="\t" when reading them. Because file_name is not prefixed with clips/, the plain audiofolder builder will not auto-attach transcripts — use the snapshot-based loader above.