Back to Language as Infrastructure
working paper2026FAC manuscript and experiment harness

Featural Acoustic Coding

Featural Acoustic Coding reframes speech recognition around acoustic and articulatory feature bundles rather than atomic phoneme labels. For N'Ko, the key observation is that tone, duration, vowel color, and onset properties should remain evidence-bearing features instead of being discarded before transcription.

Paper workspace

Live draft structure

live-draft

Artifacts

Draft PDF

Rendered FAC draft. It should remain live until feature-head experiments are reviewed.

Open artifact

Editable source

Draft PDF exists, but FAC is still theory plus harness. Read-speech evaluation is the next release gate.

Source anchors

nko-acoustic-coding/main.tex

nko-acoustic-coding/DOCUMENTATION.md

nko-acoustic-coding/experiments/READ_SPEECH_RUNBOOK.md

nko-acoustic-coding/experiments/tone_fusion_eval.py

Method tags

feature headstoneacoustic evidencephoneme composition

Ingest intersections

faccodectoneprosodyfeaturesspeech

Status

Drafted with runnable experiment harness; first read-speech evaluation pending.

Key claims

01

Phonemes are feature bundles, not indivisible classes.

02

Tone restoration is partly acoustic and cannot be solved honestly by text priors alone.

03

N'Ko is a plausible symbolic substrate for feature-grounded speech.

Public reading note

Working paper; read-speech evaluation is the next release gate.

Standard skeleton

What this paper must keep proving

Schema

problem

Phoneme labels collapse acoustic features that may matter most for low-resource tonal speech.

method

Represent speech through feature bundles before composing phonemes and rendering symbols.

implementation

Feature-head scaffolds, acoustic verifier, read-speech packet design, N'Ko feature mapping.

data

Read-speech prompts, reviewed audio, tone and feature annotations. No unreviewed output becomes truth.

evaluation

Feature prediction, tone evidence, acoustic-grounded correction ranking, and transcript faithfulness.

references

Neural speech codecs, articulatory phonology, tone modeling, acoustic representation learning.

openQuestions

Whether codec-token acoustic world models can supply the representation layer FAC needs.

Checkpoints and references

Proof chain

paperproven

FAC draft rendered

nko-acoustic-coding draft PDF

The theory and design are captured in a current render.

experimentpending

Read-speech evaluation

FAC read-speech harness

This must be completed before FAC becomes an empirical claim.

experimentpartial

Acoustic verifier signal

acoustic verifier AUC result and packet review path

Supports evidence-governance direction, not full feature-head success.

Reference links

Codec-token speech modelsextendsrelated-work slot

N'Ko substrate papersupports