New and Updated Paper Proposals — April 2026

New and Updated Paper Proposals — April 2026

Based on discoveries from the current research arc. These build on Papers 1-5 (written) and extend into new territory.

Paper 6: Updated Framing — Trajectory Attention Residuals Are a Negative Result

Original title was "Inscribing Knowledge" (blockchain provenance). The TAR experiments produced a more important story.

Title: "Trajectory Scalars Are Necessary and Sufficient: A Controlled ASR Experiment on 290K Bambara Speech Pairs"

This paper reports the full 8-way controlled experiment. The central finding is NOT that TAR improves ASR. It is that trajectory scalars alone are the essential contribution, and depth attention adds nothing.

Key results from the 291K clean run:
- N'Ko trajectory-only: 29.98
- N'Ko TAR (trajectory + depth attention): 29.95
- Latin baseline without trajectory: 100
- Trajectory scalars are mandatory for convergence on mixed data

The reproduction run currently training on 290K pairs will either confirm or update the 27.50

Additional contribution: the label contamination discovery. The bam-asr-early N'Ko column contained 130K+ Latin IPA characters. The clean_nko_text() fix improved CER by 15.4 percentage points. This is a cautionary finding for any low-resource ASR work using crowd-sourced transcriptions.

Venue: Interspeech 2026 or ACL 2026.

Paper 7: Embodied Vocabulary — Motion as Language Through N'Ko Inscription

Title: "Embodied Vocabulary: Generating N'Ko Words from Motion Dynamics via Compositional Root Assembly"

This is the paper that emerged from this session. The core claim: the same anticipation geometry (7 trajectory scalars) that improves CTC ASR decoding (Paper 4) also generates semantically meaningful N'Ko vocabulary from body movement, because the 10 inscription sigils are native Manding consonant roots.

The pipeline:
1. Body movement captured at 30Hz as 104D latent vector
2. 10 claim detectors identify dynamical events (stabilization, transition, novelty, etc.)
3. Each claim produces an N'Ko sigil that IS a Manding consonant root
4. Compound sigil sequences form pronounceable words via Manding morphological composition
5. ASR round-trip validates pronunciation (the decoder from Papers 4-5)
6. Words with sustained usage graduate into a living lexicon

Novel contributions:
- First system where physical movement generates linguistic vocabulary
- The 10 sigils were deliberately chosen as Manding consonant roots, enabling dual interpretation as dynamical labels AND morphological building blocks
- The neologism generator produced 1,974 candidate words from 36 Manding roots + 10 sigils + 8 derivational suffixes
- The vocabulary expansion targets measured gaps: technology (47 words), medicine (18), education (16) in the 2,101-entry Ankataa dictionary

Human validation: Mohamed's public N'Ko learning journey documents whether machine-generated compound words are learnable and pronounceable by a human who knows the 27-character bijective mapping.

Venue: ACL 2027 or EMNLP 2027 (needs the human data from the learning journey).

Paper 8: Word Error Rate Is Not a Valid Metric for Tonal Language ASR

Title: "Against WER: Why Character Error Rate on Bijective Scripts Is the Correct Metric for Manding ASR Evaluation"

This is a position paper. Short, sharp, argumentative. Every Bambara ASR system published today reports WER on Latin transcriptions. We argue this metric is scientifically invalid for tonal languages with inconsistent orthography.

The argument:
1. Latin Bambara has no standard spelling (multiple valid spellings per word)
2. Tone is unmarked in 98.8
3. Digraphs collapse phoneme boundaries (ny, ng, gb are single sounds written as two characters)
4. WER measures agreement with a transcriber's spelling convention, not speech understanding
5. N'Ko CER is a direct measurement of phonemic accuracy because of the bijective character-phoneme mapping
6. Formal proof: for bijective transcription function f, CER equals phoneme error rate exactly. For many-to-many f (Latin), CER is a noisy proxy.

Supporting data:
- The 27.50
- The Latin 30.32
- Latin baseline without trajectory collapses to 100

We propose N'Ko CER as the standard evaluation metric for Manding ASR, not because of cultural preference, but because it is the only metric where character-level accuracy is phonemically interpretable.

Venue: Workshop paper at Interspeech 2026 or EMNLP 2026. Short format (4 pages).

Paper 9: Compositional Generalization Across Decoders — Machine and Human

Title: "The Same Bijective Structure That Helps Machines Generalize Helps Humans Learn: Evidence from N'Ko ASR and First-Language Acquisition"

This is Paper 7's companion, focused specifically on the generalization claim. It extends Paper 5's machine results (3.65pp smaller generalization gap for N'Ko) with human learning data from Mohamed's journey.

The experimental design:
- Machine arm: Paper 5's compositional generalization results (N'Ko decoder generalizes better to unseen words than Latin decoder, 37.81pp vs 41.46pp gap)
- Human arm: Mohamed learns N'Ko characters, then attempts to pronounce unseen words. Accuracy measured via NKoScribe ASR. Compare against a control group learning Latin Bambara spelling.
- Bridge: the same property (bijective encoding) explains both results. The machine doesn't need the word in training data. The human doesn't need the word in their vocabulary.

This paper needs the human data to be collected, so it is the longest-horizon proposal. But the machine half is already proven, and the experimental design is clean.

Venue: CogSci 2027 or FAccT 2027.

Updated Papers 4 and 5

Paper 4 (Script Design Affects ASR) needs:
- Label contamination disclosure paragraph
- Updated data count (290K instead of 297K, document why)
- Confirmed reproduction CER (from the current Vast.ai run)

Paper 5 (Deployment Properties) needs:
- Aligned data count with Paper 4
- Djoko integration results (if Djoko training completes)

Summary

Total: 9 papers. 5 written, 1 needs update, 3 new. Papers 1-3 and 8 can submit immediately. Papers 4-6 gate on the reproduction run. Papers 7 and 9 gate on the vocabulary engine and human learning data.