N'Ko as Computational Infrastructure
This paper treats N'Ko as computational infrastructure rather than an output font. It consolidates script-native ASR, phonemically interpretable evaluation, and governance around tone correction into one measurement thesis: the script used for evaluation changes what an error rate means.
Paper workspace
Live draft structure
Artifacts
Draft PDF
Rendered from the current N'Ko ASR consolidation draft. Public for reading, not final citation copy.
Open artifactSplit paper: script invisibility
Companion split-paper render from the final N'Ko paper set.
Open artifactSplit paper: phonemic evaluation
Companion split-paper render for N'Ko phonemic evaluation.
Open artifactSplit paper: script-native ASR anchor
Companion split-paper render for the direct script-native ASR anchor.
Open artifactSplit paper: AGP deployment
Companion split-paper render for governed deployment and correction.
Open artifactEditable source
Live manuscript source exists and should continue changing until the final release gate is explicit.
Source anchors
nko-brain-scanner/paper/current/paper_canonical_nko_agp_20cer.tex
nko-brain-scanner/paper/final/01-script-invisibility/paper.tex
nko-brain-scanner/paper/final/02-phonemic-evaluation/paper.tex
nko-brain-scanner/paper/final/03-script-native-asr-anchor/paper.tex
nko-brain-scanner/paper/final/04-agp-deployment/paper.tex
Method tags
Ingest intersections
Status
31 pages, builds clean. Companion papers in preparation.
Key claims
01
Character error rate is not script-neutral for low-resource languages.
02
N'Ko can expose phonemic structure that Latin-script intermediaries hide.
03
Tone correction needs acoustic evidence and governance, not text fluency alone.
Public reading note
Readable public summary now; full PDF should wait for release decision.
Standard skeleton
What this paper must keep proving
problem
Standard ASR metrics treat script choice as neutral, then hide phonemic structure inside an inherited orthography.
method
Train and evaluate script-native N'Ko recognition, then separate proposal generation from acoustic governance.
implementation
Whisper encoder, CoreML CTC head, deterministic candidate generation, Swift ranker, acoustic verification gates.
data
Private training/evaluation runs plus governed live packets. Public corpora remain broad acoustic candidates, not live truth.
evaluation
CER, phonemic interpretation, deployment gates, and acoustic verifier behavior. Live ASR remains a separate proof.
references
Low-resource ASR, CTC speech recognition, script-aware NLP, tone restoration, governed correction loops.
openQuestions
How much of the remaining error is acoustic evidence, script modeling, tone recovery, or live-mic frontend mismatch.
Checkpoints and references
Proof chain
Offline script-native ASR result
nko-brain-scanner training/evaluation artifacts
Supports the offline recognition claim. It does not prove stable live iPhone ASR.
On-device deployment path
Whisper/CoreML encoder plus CTC head iPhone harness
Proves the pipeline can execute on device, with live reliability still gated separately.
Governed correction boundary
acoustic-gate packet review and no-machine-truth gates
Prevents garbage output and public labels from silently becoming truth.
Reference links
External CTC and low-resource ASR referencesneeds-citationrelated-work slot