ayush-shunyalabs/Indic_ASR_Eval

Indic ASR Eval

A curated evaluation set for Indic-language automatic speech recognition. 100 samples are sampled (seed = 42) from each (source dataset × language) cell of seven public Indic ASR corpora. Each source corpus is published as its own dataset config with a single test split, at 16 kHz.

• Rows: 6,169 across 7 configs
• Total audio: ~13.3 hours
• Sampling rate: 16 kHz (mono)
• Split: test (single split in every config)

Configs

Config	Rows	Notes
kathbath	1,200	AI4Bharat Kathbath (read speech, 12 languages)
kathbath_noisy	1,200	Noisy variant of Kathbath
fleurs	1,100	Google FLEURS (11 Indic languages)
indictts	900	AI4Bharat IndicTTS test utterances
respin	900	IISc RESPIN test (including Bhojpuri, Chhattisgarhi, Magahi, Maithili)
commonvoice	769	Mozilla Common Voice v11 test sets
gramvaani	100	Gram Vaani 3-hour Hindi evaluation

Malayalam has 69 commonvoice samples (rather than 100) because that is the total available in the CV v11 Malayalam test set.

Languages

Bengali · Bhojpuri · Chhattisgarhi · Gujarati · Hindi · Kannada · Magahi · Maithili · Malayalam · Marathi · Odia · Punjabi · Sanskrit · Tamil · Telugu · Urdu

Row totals across all configs:

Language	Samples
Hindi	700
Bengali	600
Marathi	600
Kannada	500
Odia	500
Tamil	500
Telugu	500
Malayalam	469
Gujarati	400
Punjabi	400
Urdu	400
Sanskrit	200
Bhojpuri	100
Chhattisgarhi	100
Magahi	100
Maithili	100

Schema (all configs)

Field	Type	Description
audio	Audio(sampling_rate=16_000)	Mono waveform, decoded lazily
transcript	string	Reference transcription (native script)
language	string	Title-case language name (e.g. Hindi)
duration	float32	Clip duration in seconds
dataset	string	Source dataset id (matches the config name)

Usage

Load a single source

from datasets import load_dataset

fleurs = load_dataset("ayush-shunyalabs/Indic_ASR_Eval", "fleurs", split="test")
print(fleurs[0]["language"], fleurs[0]["transcript"])

Load all configs and concatenate

from datasets import load_dataset, concatenate_datasets

CONFIGS = [
    "commonvoice", "fleurs", "gramvaani",
    "indictts", "kathbath", "kathbath_noisy", "respin",
]
all_ds = concatenate_datasets([
    load_dataset("ayush-shunyalabs/Indic_ASR_Eval", c, split="test")
    for c in CONFIGS
])

Filter by language

hindi = fleurs.filter(lambda r: r["language"] == "Hindi")

Evaluate an ASR model

from datasets import load_dataset
from akshara_mulyankan import evaluate_pairs

ds = load_dataset("ayush-shunyalabs/Indic_ASR_Eval", "kathbath", split="test")
hindi = ds.filter(lambda r: r["language"] == "Hindi")

preds = [my_asr_model(r["audio"]["array"], sr=16000) for r in hindi]
refs = hindi["transcript"]

result = evaluate_pairs(list(zip(refs, preds)), language="hindi")
print(result["summary"])

Licensing & attribution

Each sample inherits the license of its source dataset — please consult the upstream corpus before redistribution:

• Kathbath / Kathbath-Noisy / IndicTTS: AI4Bharat (CC-BY-4.0)
• FLEURS: Google Research (CC-BY-4.0)
• Common Voice v11: Mozilla Foundation (CC0-1.0)
• RESPIN: IISc Bangalore
• Gram Vaani: Gram Vaani

This compilation itself is released under CC-BY-4.0. If you use it, please also cite the upstream datasets above.