CoSHE-500 — Conversational Hinglish Speech Evaluation

CoSHE-500 is a 500-clip evaluation set of conversational Hindi–English code-switched (Hinglish) speech, with verbatim mixed-script (Devanagari + Latin) reference transcriptions. It is designed to measure real-world code-switch ASR quality — the setting where most Hindi/English models degrade.

Source. CoSHE-500 is a curated 500-clip subset of CoSHE — soketlabs/CoSHE-Eval by Soket Labs — redistributed under its original CC-BY-NC-4.0 license (attribution, non-commercial). All credit for the underlying audio and transcriptions goes to the CoSHE / Soket Labs team. We only sampled, chunked to ~30 s, and tidied the columns.

• Clips: 500 conversational utterances (~30 s each)
• Split: eval
• Columns: audio (16 kHz mono), transcription (verbatim, mixed-script)
• Languages: Hindi + English, code-switched within utterances
• Script: Devanagari for Hindi, Latin for English — kept verbatim (no forced transliteration)

Usage

from datasets import load_dataset
ds = load_dataset("Trelis/CoSHE-500", split="eval")
print(ds[0]["transcription"])
# ds[0]["audio"]  -> {"array": np.ndarray, "sampling_rate": 16000}

Recommended scoring

Hinglish WER is sensitive to the normaliser. We recommend a script-safe Indic normaliser that:

• applies Unicode NFC + Indic normalisation, keeping Devanagari matras and nuktas (ड ≠ ड़),
• lowercases Latin,
• strips punctuation (including the apostrophe), keeps letters/marks/numbers,
• does not transliterate and does not number-normalise.

The default Whisper text normaliser strips Devanagari matras and inflates Devanagari WER — do not use it for this benchmark. Report corpus WER (sum of edits / sum of reference words), lower is better.

Note for code-switch scoring: systems that emit English loanwords in Latin or numbers as digits will be penalised against these Devanagari references by a translit-blind WER. State your output convention when reporting.

Intended use

Evaluation / benchmarking of ASR systems on conversational Hinglish. Not a training set. Please do not train on CoSHE-500 — keep it held out so it remains a fair public yardstick.

License & attribution

CC-BY-NC-4.0, inherited from the source dataset soketlabs/CoSHE-Eval (Soket Labs). Attribution required; non-commercial use only. Please cite CoSHE / Soket Labs when you use this benchmark. This repackaging (subset, 30 s chunking, column cleanup) is provided for convenient evaluation; the audio and reference transcriptions remain the property of the original CoSHE authors.