kriti

<div align="center">

kriti 🎙️

a compact acoustic punctuation architecture for nepali asr

kriti couples a nepali-only rnnt transcription graph with a 1,025-parameter acoustic terminal-danda head. the released live graph contains exactly 119,462,146 parameters.

source · architecture · training · benchmark · data · references

</div>

the contribution

kriti turns a multilingual hybrid checkpoint into a focused nepali inference graph and adds an acoustic punctuation branch that shares the speech encoder. the architecture has four released contributions:

1. a nepali-only rnnt path with the ctc branch and 21 additional language heads removed
2. a runtime-compacted prediction embedding with 257 nepali output rows
3. a 1,025-parameter acoustic head for terminal devanagari danda recovery
4. content-bound loading that verifies artifact hashes, graph shape, parameter count, and prediction identity

the pretrained conformer encoder and nepali rnnt weights come from the mit-licensed ai4bharat nepali indicconformer at revision cd09ba7720f3b17d259f6bfd03e1463bc5ba517d. kriti retains those upstream weights, applies deterministic deployment pruning, and trains the acoustic danda head. the full technical lineage is recorded in references.md and notice.

architecture

16 khz audio → log-mel → conformer encoder → nepali rnnt → transcript
                                  │
                                  └→ pooled encoder states → danda head → final text

the transcription graph contains 119,461,121 live parameters. the punctuation branch pools the same encoder states into 1,024 mean and standard-deviation features, then applies one logistic layer and a frozen threshold of 0.711. the combined graph reaches 119,462,146 live parameters and uses one acoustic network.

the full layer dimensions, pruning contract, artifact hashes, and loader invariants are in the architecture note.

measured result

the current evidence is a frozen 3,630-utterance development snapshot containing 19 fully evaluated open systems. every included system completed two fresh-load replicates with matching prediction hashes.

pi wer means punctuation-insensitive word error rate. exact primary error ratios share a competition rank, with raw wer reported as a separate descriptive measure. kriti shares exact rank 1 with the official ai4bharat nepali rnnt and records the lowest raw wer within the evaluated snapshot.

the snapshot uses a development view that also guided kriti model selection. the planned campaign ended after 19 complete systems, and the published table contains results only for that completed field. stronger claims require a separately frozen evaluation covering real microphones, streaming behavior, domain shift, latency, and product punctuation.

the complete machine-readable record is benchmark.json. it contains all 19 systems, immutable revisions, decoder settings, exact metrics, and replicate hashes.

data record

the punctuation-head recipe and benchmark use the ne-commercial-v1 profile: 402,905 accepted clips and 630.2761 decoded hours, including 393,002 training clips and 608.9182 training hours.

each source retains its license, access terms, and attribution requirements. audio, transcripts, speaker metadata, and row manifests stay with their licensed data environment. the complete record is in the data note.

quick start

git clone https://github.com/harrrshall/kriti
cd kriti
python -m venv .venv
source .venv/bin/activate
pip install -e '.[runtime]'
kriti audio.wav

from kriti import load_model

model = load_model()
print(model.transcribe("audio.wav"))

the loader fetches the public artifacts, verifies both sha-256 values, builds the compact nepali graph, checks the exact live parameter count, and runs rnnt decoding with terminal danda recovery.

license and credit

kriti source code is released under mit. the released weights preserve the ai4bharat indicconformer lineage and its mit terms. datasets and benchmark checkpoints remain governed by their respective authors, licenses, model cards, and access conditions. notice and references.md provide the attribution record.