FAC Read-Speech Evaluation Harness
The FAC harness exists to test whether feature-level acoustic cues can be gathered and evaluated without confusing prompted audio, transcript truth, and training labels. It is deliberately fail-closed until reviewed labels exist.
Paper workspace
Live draft structure
Artifacts
Experiment runbook
The harness is source and runbook first. No PDF is needed until a real read-speech result exists.
source-only
Editable source
Experiment runbook and code exist. It remains blocked on real read speech aligned to known toned N'Ko text.
Source anchors
nko-acoustic-coding/experiments/READ_SPEECH_RUNBOOK.md
nko-acoustic-coding/experiments/run_read_speech_eval.py
nko-acoustic-coding/Makefile
nko-acoustic-coding/experiments/tone_fusion_eval.py
nko-acoustic-coding/experiments/RESULTS.md
Method tags
Ingest intersections
Status
Experiment scaffolded; not yet enough reviewed live labels for a correctness claim.
Key claims
01
Feature supervision must be collected separately from transcript guesses.
02
Prompted family audio can support coverage, not automatic correctness.
03
FAC should enter the model only after replayable reviewed evidence exists.
Public reading note
Public summary only; no accuracy claim yet.
Standard skeleton
What this paper must keep proving
problem
FAC needs a decisive acoustic tone evaluation, but lesson commentary cannot supply aligned tone ground truth.
method
Require a read WAV aligned to known tone-marked N'Ko text, then score tone-diacritic evidence instead of text fluency.
implementation
Makefile read-speech target, read-speech runbook, tone fusion evaluator, and gold-tone extraction logic.
data
Admissible data is read speech of known toned text. Lesson audio and OCR are useful context but not aligned acoustic truth.
evaluation
Tone-diacritic error rate and feature/tone fidelity after alignment.
references
Tone language ASR, pitch tracking, tone diacritic evaluation, read-speech corpus design.
openQuestions
Collecting the first small, clean read-speech packet with known toned N'Ko text.
Checkpoints and references
Proof chain
Harness source exists
nko-acoustic-coding read-speech target
The code path for running the evaluation exists.
Real acoustic TDER
aligned read-speech WAV plus gold N'Ko text
The decisive metric is intentionally blocked until the right data exists.