Aarjan/kriti-telephony

Kriti Telephony

A Nepali speech-recognition model tuned for real phone-call audio: call-center conversation, code-switching, and narrowband lines. It is a domain-adapted version of Kriti by Naamche Labs.

Full methodology, benchmarks, and a reproducible evaluation harness live in the repository: github.com/Naamche-Labs/kriti-telephony.

Results

All numbers are word error rate (WER), lower is better, scored with the official NepTel scorer on its current references. General Nepali is a held-out clean-read set with human references, reported alongside as a check against overfitting.

Model	NepTel WER (telephony)	General Nepali WER
Kriti (base)	40.6	4.1
Kriti Telephony	32.4	6.3

On NepTel, Kriti Telephony posts the lowest WER (32.4 vs the previous best 33.8), and it stays strong on general Nepali at 6.3, close to base Kriti. It is good on both: strong on phone audio and still capable on clean audio. For pure studio or read-speech work, base Kriti has a small edge; for anything involving calls, use Kriti Telephony.

Usage

from huggingface_hub import hf_hub_download
import nemo.collections.asr as nemo_asr

path = hf_hub_download("Aarjan/kriti-telephony", "kriti_telephony.nemo")
model = nemo_asr.models.ASRModel.restore_from(path, strict=False)
model.cur_decoder = "rnnt"
print(model.transcribe(["call.wav"], language_id="ne"))

Kriti runs on the AI4Bharat NeMo fork and, on Hopper GPUs (H100), needs NUMBA_CUDA_USE_NVIDIA_BINDING=1. Exact package pins are in the repository's ENVIRONMENT.md.

How it was built

Kriti Telephony was produced by knowledge distillation: a strong existing Nepali model relabels conversational Nepali audio in the target transcription style, Kriti is fine-tuned on that, and a large general-Nepali anchor with human labels is mixed in to preserve general ability. Full detail and the teacher disclosure are in the repository's METHODOLOGY.md.

Limitations

• Telephony-tuned. Base Kriti still has a small edge on clean audio (4.1 vs 6.3).
• NepTel is a single benchmark (3 calls, one vendor); the lead over the previous best (32.4 vs 33.8) is a statistical tie by paired bootstrap. Treat the direction as the finding, not the exact decimal.
• Part of the telephony gain reflects matching a shared reference transcription style; the general-Nepali number is reported to bound this, and a neutral benchmark is the planned next validation.

License

MIT. Third-party datasets, benchmarks, and models retain their own licenses; see the repository for full attribution.