SkunkWorkLabs/hindi-asr-benchmark

SkunkWorks Hindi STT Benchmark

Hindi ASR benchmark evaluating SkunkWorks alongside major commercial Hindi STT providers (ElevenLabs, Deepgram, Sarvam) across 6 held-out evaluation subsets.

Subsets

config	source	n
kathbath	AI4Bharat Kathbath	1,929
kathbath_noisy	Kathbath noisy mic conditions	1,929
commonvoice	Mozilla Common Voice Hindi	1,727
mucs	MUCS 2021 Hindi subtask	3,897
fleurs	Google FLEURS hi_in test	418
indictts	AI4Bharat IndicTTS	100

Total: ~10,000 utterances, ~15.5 hours of held-out Hindi audio.

Systems compared

• skunkworks — our finetuned NeMo Conformer-RNNT (0.6B params, ITN-style output)
• elevenlabs — ElevenLabs Scribe v1
• deepgram — Deepgram Nova-2
• sarvam — Sarvam Saarika

Schema (per row)

audio:                  Audio        # 16 kHz mono, bytes embedded
transcript:             str          # human reference (spelled-out Devanagari)
transcript_normalized:  str          # same — original benchmark normalization
ref_norm:               str          # Vistaar + digit-expansion normalized ref
duration:               float
id:                     str
subset:                 str
language:               str  ("hi")

skunkworks:             str          # our model raw output (ITN style)
skunkworks_norm:        str          # Vistaar-normalized

elevenlabs:             str
elevenlabs_normalized:  str

deepgram:               str
deepgram_normalized:    str

sarvam:                 str
sarvam_normalized:      str

Evaluation methodology

Following AI4Bharat's Vistaar / IndicWhisper paper (Bhogale et al., Interspeech 2023), we apply the following normalization to both refs and hyps before computing WER/CER:

1. Strip ASCII + Indic punctuation (string.punctuation plus ।۔'-॥ and ZWJ/ZWNJ).
2. Expand ASCII digit runs and 1st/2nd/3rd ordinals to spelled-out Hindi (necessary to compare ITN-style outputs against spelled-out refs).
3. IndicNormalizerFactory("hi").normalize(s) (NFC → NFD).
4. Collapse multiple spaces.

The ref_norm and <system>_norm columns expose this normalization for reproducibility.

Headline results (Vistaar-normalized WER %)

Subset	skunkworks	elevenlabs	deepgram	sarvam
indictts	9.75	13.20	15.41	14.71
fleurs	17.29	11.93	21.22	15.74
kathbath	16.82	13.32	20.55	16.62
kathbath_noisy	19.06	13.16	21.98	17.75
commonvoice	24.16	17.02	28.34	19.32
mucs	24.60	10.97	20.54	12.72

See results/comparison.md and results/eval_ringg_summary.json for the full breakdown including raw WER/CER and per-utterance pairs.

Usage

from datasets import load_dataset
ds = load_dataset("SkunkWorkLabs/hindi-asr-benchmark", "kathbath", split="eval")
print(ds[0]["transcript"])
print(ds[0]["skunkworks"])     # our model
print(ds[0]["elevenlabs"])     # competitor

License

Audio files retain their original source licenses (Common Voice = CC-0, FLEURS = CC-BY-4.0, Kathbath/IndicTTS = AI4Bharat licenses, MUCS = Microsoft research). Predictions and metadata in this benchmark are released under Apache 2.0.