NepTel v0.1 — Nepali real-telephony ASR benchmark

75 scored segments / 2,375 reference words of real Nepali call-center audio (genuine two-party customer-support calls), with human-reviewed reference transcripts. To our knowledge this is the first public Nepali ASR benchmark on real call audio rather than read-aloud speech.

This repository exists so anyone can benchmark a Nepali ASR system without any access request: the audio is cut and ready, no gate, no approval step.

Evaluation only. Do not train on this data. Canary: NEPTEL-CANARY-2026-8f3a1c92

Contents

• audio/ — 77 segment wavs (22.05 kHz mono, ~31 minutes total)
• references.json — reference transcripts and per-segment provenance

references.json lists all 77 segments; the 2 carrying an excluded field (one caught by the speaking-rate gate, one flagged unintelligible in native-speaker review) are skipped by the scorer, leaving the 75 scored segments / 2,375 words quoted in published results. They are kept rather than deleted so the exclusions stay auditable.

Score a system

git clone https://github.com/Ampixa/nepaliconformer
cd nepaliconformer/benchmark
python fetch_audio.py neptel_audio        # pulls this dataset, no login

# then plug your system in — pick the adapter that fits:
python run_system.py --hf openai/whisper-large-v3 --name whisper --lang ne
python run_system.py --nemo your_model.nemo --name mine
python run_system.py --cmd "your-cli -f {audio}" --name mine
python run_system.py --py yourpkg.module:transcribe --name mine

Each writes outputs/<name>.json and prints the WER; add --compare outputs/nepali-conformer-offline.json for a paired bootstrap against ours.

The scorer normalizes both sides identically (digits to spoken Nepali words, punctuation stripped) and reports a paired bootstrap when given --compare.

Current standings

system	WER
nepaliconformer offline (121M)	33.81
Kriti — Naamche Labs (119M)	40.59
nepaliconformer streaming (520 ms)	59.87
MMS-1B-all — Meta, zero-shot (965M)	81.01
IndicWav2Vec-Nepali — community mirror (94M)	86.57
Whisper-large-v3 — zero-shot, tuned (809M)	96.29

Every system's raw hypotheses are published in the repo, so all of these are re-derivable. We would love to be beaten — run the scorer and open a PR.

Reference protocol

1. Audio cut into fixed 25-second windows from the vendor's dual-channel per-speaker files.
2. Reference drafts by Google Chirp 2 (ne-NP).
3. A native Nepali speaker reviewed every segment against the audio: accept / correct / flag. Flagged segments are excluded. 95.9% of usable segments in the most recent batch were accepted verbatim.
4. A speaking-rate gate (>6 words/sec) guards against transcription-engine hallucinations becoming ground truth.

Known limitations: 3 calls from one vendor; vendor-side PII muting leaves digital-zero gaps in some segments; references are reviewed drafts rather than from-scratch transcriptions; systems trained on Chirp 2 pseudo-labels (including ours) share label lineage with the reference drafts. Sized for ±2-point deltas between systems, not for absolute-truth WERs. Full provenance, including declared conflicts of interest, is in PROVENANCE.md.

Source audio and license

Segments are cut from InfoBayAI/Nepali_Call_Center_Audio_Dataset_Dual_Channel, published by InfoBayAI under CC-BY-4.0, which permits redistribution with attribution. The audio here is that dataset, cut into segments; the reference transcripts are ours, also CC-BY-4.0. Please credit both InfoBayAI (source audio) and Ampixa Labs (segmentation and references).