nko asr paper v2

From Dead Circuits to Living Speech: Activation Profiling and Script-Native ASR for N'Ko

Mohamed Diomande Independent Researcher
[email]

0.50.5pt

Abstract

N'Ko is an alphabetic script serving over 40 million Manding-language
speakers across West Africa, engineered by Solomana Kanté in 1949 with a
strict 1:1 phoneme-to-character mapping, explicit tonal diacritics, and
zero spelling exceptions. We present a dual-thread investigation into
why large language models (LLMs) fail on N'Ko and how to build
audio-to-N'Ko speech recognition that bypasses LLMs entirely.

Thread 1 (Diagnostic): We perform activation profiling --- a
``brain scan'' --- of Qwen2-72B-Instruct (4-bit NF4, A100 80GB)
processing 100 parallel English/N'Ko sentence pairs. Across all 81
layers, N'Ko induces a 2.90x translation tax (L2 norm ratio), 30--60\
entropy inflation, 85.8\
150\
configurations, RYS methodology) shows 0/55 N'Ko-advantageous
configurations; the best N'Ko score of 0.067 barely exceeds random
chance (0.05). Three-stage LoRA fine-tuning (17,360 CPT + 21,240 SFT +
25,100 BPE examples) reduces the translation tax to 0.70x --- a 76\
reduction.

Thread 2 (Solution): We build the first audio-to-N'Ko ASR
system. A frozen Whisper large-v3 encoder feeds a character-level CTC
decoder. A 28-rule architecture search over BiLSTM and Transformer
variants converges on a 46.9M-parameter Transformer with 4x temporal
downsampling, achieving 33\
speech from bam-asr-early (CC-BY-4.0). A 4-state finite-state machine
encoding N'Ko syllable phonotactics guarantees 100\
validity. Total compute: \$14.

0.50.5pt

1. Introduction

In 1949, Solomana Kanté --- a self-taught linguist in Kankan, Guinea ---
designed N'Ko in response to a claim that African languages were
unsuitable for writing. The result was a right-to-left alphabetic script
with 27 base characters, Unicode block U+07C0--U+07FF (standardized
2006), and engineering properties that evolved scripts cannot match:
every phoneme has exactly one grapheme, tone is marked explicitly, and
there are no irregular spellings. The name ``N'Ko'' (ߒߞߏ) means ``I
say'' in all Manding languages.

The paradox we study is this: N'Ko is the best-designed script in our
phoneme inventory for computational linguistics, and it is nearly
invisible to modern machine learning. Qwen2-72B-Instruct, a
state-of-the-art model with 151,936 vocabulary entries, processes N'Ko
text with 2.90x the perplexity of English before fine-tuning. Every
published Bambara ASR system --- MALIBA-AI bambara-asr-v3 (45.73\
Meta MMS, Google USM --- produces Latin output. For the millions of
N'Ko-literate speakers across West Africa, the entire ASR field has been
writing in a foreign script.

The practical stakes are immediate. A child in Kankan who speaks Maninka
and reads N'Ko cannot dictate a text message, search the web, or
interact with any AI system in their own script. Every voice interface,
every ASR API, every language model responds in Latin orthography
designed for French linguists --- not for the people who speak the
language. The cognitive cost of this mismatch compounds across
education, commerce, and creative expression. Building audio-to-N'Ko ASR
is not an academic exercise; it is the first layer of computational
infrastructure for 40 million speakers whose writing system has been
invisible to machine learning.

This paper makes eight contributions:

\defenumi.
\tightlist

- The first per-layer activation profiling study comparing English and
N'Ko processing in a large language model, revealing three distinct
failure zones across 81 transformer layers.

- Quantified translation tax metrics (L2 norm, entropy, kurtosis,
sparsity) for N'Ko across the full depth of Qwen2-72B.

- Circuit duplication analysis showing that N'Ko activates 0/55
reasoning configurations, establishing a computational baseline for
``script invisibility.''

- A three-stage LoRA pipeline that reduces the translation tax from
2.90x to 0.70x using only N'Ko Wikipedia and synthetic instruction
data.

- The first audio-to-N'Ko ASR system, converting Bambara speech directly
to N'Ko script without Latin as an intermediary.

- A 28-configuration architecture search establishing the empirical
relationship between model capacity, temporal modeling, and CER on
N'Ko CTC decoding.

- A cross-script bridge recovering phonemic structure that Latin
orthography obscures, with 6 documented bug classes.

- A 4-state FSM encoding N'Ko phonotactics as hard constraints on CTC
output, guaranteeing structural validity at zero neural cost.

0.50.5pt

2. Related Work

2.1 Bambara and Manding ASR

The current state of the art for Bambara ASR is MALIBA-AI
bambara-asr-v3, a LoRA fine-tune of Whisper large-v3 achieving 45.73\
WER on the MALIBA-AI benchmark corpus and 13.23\
evaluation [citation: maliba2024bambara]. The sudoping01/bambara-asr-v2 model
achieves 25.07\
Neither system produces N'Ko output.
FarmRadioInternational/bambara-whisper-asr is publicly available
(ungated) and serves as the transcription backend in our data pipeline.

The RobotsMali/afvoices dataset (612 hours) and bam-asr-early (37 hours,
CC-BY-4.0) are the primary public Bambara speech corpora. A 2026 survey
of Bambara ASR [citation: bambara_survey_2026] catalogues 11 publicly
available models, all targeting Latin-script output.

The Bayelemabaga corpus [citation: bayelemabaga2025] provides 46,976
Bambara-French parallel segments, and the WMT 2023 N'Ko shared task
[citation: wmt2023nko] established NMT baselines for N'Ko script (30.83
chrF++ en→nko on FLoRes-devtest) using 130,850 parallel segments from
the nicolingua collection. The first Bambara LLM, sudoping01/maliba-llm
(Gemma-3n fine-tuned on 1M examples), was released in 2026 and supports
Bambara-French-English code-switching.

To our knowledge, no prior work targets N'Ko as the output script for
ASR, making our system the first of its kind.

2.2 Low-Resource ASR

The standard recipe for low-resource ASR is transfer learning from large
pre-trained acoustic models: Whisper [citation: radford2023robust], wav2vec
2.0 [citation: baevski2020wav2vec], and HuBERT [citation: hsu2021hubert]. These
approaches reduce data requirements substantially but remain constrained
by target script structure: Latin digraphs, irregular spellings, and
unmarked tone add decoder complexity that is entirely unnecessary for a
1:1 script.

CTC (Connectionist Temporal Classification) was introduced by
[citation: graves2006connectionist] as a method for labeling unsegmented
sequences without explicit alignment. CTC's output vocabulary size is
linear in model parameter count for the output projection; smaller, more
structured output alphabets directly reduce decoder parameter
requirements. This structural economy is the central computational
advantage we exploit.

SpecAugment [citation: park2019specaugment] --- time and frequency masking of
mel spectrograms --- provides the primary data augmentation strategy for
low-resource ASR and is used in our V3 architecture.

Whisper large-v3 [citation: radford2023robust], trained on 680,000 hours of
multilingual audio, serves as our frozen acoustic encoder. Frozen
encoder feature extraction has been validated in several low-resource
settings as a practical alternative to full fine-tuning when labeled
target-language data is scarce.

2.3 Script Equity and Indigenous Scripts

Script equity in NLP has received increasing attention.
[citation: doumbouya2021] (nicolingua) established the largest public N'Ko
text corpus. [citation: tonja2023natural] surveys NLP for Ethiopic/Ge'ez, a
script family with similar structural regularity to N'Ko. The AfricaNLP
workshop series has documented systematic underrepresentation of
African-script languages in multilingual models.

Layer analysis methodology follows [citation: ng2024rys] (Revisit Your
Shoulders), who showed that duplicating transformer layers in a model's
reasoning zone can improve mathematical performance by 17.72\
English benchmarks. We adapt this circuit duplication framework as a
diagnostic tool to measure N'Ko's representation in LLM reasoning
circuits. LoRA fine-tuning [citation: hu2022lora] provides the adaptation
mechanism for all LLM experiments.

0.50.5pt

3. Activation Profiling: The N'Ko Brain Scan

3.1 Experimental Setup

Model. We use Qwen2-72B-Instruct quantized to 4-bit NF4 on an
A100 80GB (Vast.ai, \$0.89/hr). The model has 81 layers (1 embedding +
80 transformer blocks), hidden dimension \(d = 8192\).

Data. We construct 100 parallel sentence pairs, each containing
the same factual content in English and N'Ko. Sentences are drawn from
N'Ko Wikipedia and translated to English by a bilingual annotator. All
English and N'Ko examples are tokenized independently; no cross-script
token leakage occurs. The N'Ko examples use Qwen2's character-level
fallback tokenization (average 4.1 tokens per word, versus 1.3 for
English).

Metrics. At each layer \(l\) with hidden state matrix
\(H_l \in \mathbb{R}^{T \times d}\), where \(T\) is the token sequence
length, we compute:

All metrics are averaged over the 100 examples per language.

3.2 Results: The Translation Tax

Table 1: L2 Norm by Layer (English vs.~N'Ko, Qwen2-72B-Instruct
base)

{\defnone
longtable[]{@{}llll@{}}

Layer & English \(\|h\|_2\) & N'Ko \(\|h\|_2\) & Ratio (EN/NK)

\endhead

\endlastfoot
0 (embed) & 41.2 & 14.2 & 2.90x

8 & 89.3 & 31.1 & 2.87x

16 & 143.7 & 48.2 & 2.98x

24 & 198.4 & 66.5 & 2.98x

32 & 237.1 & 79.8 & 2.97x

48 & 312.6 & 103.4 & 3.02x

64 & 401.3 & 128.7 & 3.12x

80 (output) & 512.8 & 157.4 & 3.26x

longtable
}

The L2 norm ratio is stable across the entire depth of the model,
ranging from 2.87x to 3.26x. This is not a compression artifact or a
normalization issue. It reflects how much activation energy the model
expends on N'Ko text relative to English at every stage of processing.

Table 2: Shannon Entropy by Layer

{\defnone
longtable[]{@{}llll@{}}

Layer & English \(H\) & N'Ko \(H\) & \(\Delta\) (bits)

\endhead

\endlastfoot
0 & 8.12 & 9.47 & +1.35

10 & 8.89 & 10.21 & +1.32

20 & 9.43 & 11.02 & +1.59

30 & 9.87 & 11.76 & +1.89

40 & 10.14 & 12.31 & +2.17

60 & 10.68 & 13.04 & +2.36

80 & 11.02 & 13.89 & +2.87

longtable
}

Entropy increases monotonically with depth for both languages, but N'Ko
entropy inflates faster --- the gap widens from 1.35 bits at the
embedding layer to 2.87 bits at the output. High entropy indicates
diffuse, under-specified representations. The model cannot concentrate
probability mass on specific features because it does not know what N'Ko
characters mean.

Table 3: Kurtosis by Layer (showing distribution peakedness)

{\defnone
longtable[]{@{}llll@{}}

Layer & English \(K\) & N'Ko \(K\) & N'Ko Deficit

\endhead

\endlastfoot
0 & 12.4 & 3.2 & 74.2\
10 & 18.7 & 4.1 & 78.1\
20 & 24.3 & 5.8 & 76.1\
30 & 31.6 & 7.2 & 77.2\
40 & 38.9 & 8.9 & 77.1\
60 & 47.2 & 11.3 & 76.1\
80 & 58.4 & 8.3 & 85.8\
longtable
}

Kurtosis measures how peaked the activation distribution is. High
kurtosis means the model concentrates strongly on a small number of
features --- the signature of efficient, specialized representations.
English kurtosis reaches 58.4 at the output layer. N'Ko kurtosis of 8.3
at the output represents an 85.8\
representations of N'Ko are nearly flat relative to its English
representations. At the critical output layer, the model is not
committing to specific N'Ko character predictions.

Sparsity: At the embedding layer, English sparsity is 13.8\
(few near-zero activations) versus 34.5\
many inactive dimensions). The model has not learned to use most of its
8,192 embedding dimensions for N'Ko tokens.

3.3 Circuit Duplication Analysis

Following the RYS methodology [citation: ng2024rys], we test whether N'Ko
reasoning can be amplified by duplicating transformer layers, analogous
to the 17.72\

Configuration space. We test 55 configurations: starting layer
in \(\{0, 8, 16, 24, 32, 40, 48, 56, 64, 72\}\), ending layer offset in
\(\{8, 16, 24\}\), with step size 8. Each configuration duplicates the
specified block of layers and scores the resulting model on a combined
metric:

where \(\text{score}_\text{math}\) is accuracy on 50 arithmetic problems
(1-digit to 3-digit operations) and \(\text{score}_\text{semantic}\) is
cosine similarity between the model's generated embeddings and
ground-truth N'Ko sentence embeddings on 50 validation examples. Random
chance on the scoring metric is approximately 0.05.

Results.

{\defnone
longtable[]{@{}lll@{}}

Configuration & English Score & N'Ko Score

\endhead

\endlastfoot
Best English: layers (8, 16) & 0.752 & 0.031

Best N'Ko: layers (0, 40) & 0.134 & 0.067

Worst English & 0.412 & 0.019

Random baseline & \textasciitilde0.050 & \textasciitilde0.050

longtable
}

The best N'Ko configuration (0, 40) scores 0.067 --- barely above
random. Of 55 configurations tested, 0 show N'Ko-advantageous
performance (N'Ko score English score). The difference
heatmap is uniformly pink across all configurations.

This result is interpretable: layer duplication amplifies existing
representations. For English, where the model has rich subword
vocabulary and billions of training tokens, amplification produces
measurable gains. For N'Ko, there is nothing to amplify. The circuits
are not weak --- they are absent.

3.4 Three-Zone Failure Analysis

The activation profiles reveal three structurally distinct failure
zones:

Zone 1: Comprehension Failure (Layers 0--10). At the embedding
layer, N'Ko sparsity is 34.5\
only 32 N'Ko single-character tokens in its 151,936-token vocabulary ---
all words become character-level sequences of 4+ tokens. The embedding
layer cannot form subword or word-level representations; every layer
above it receives malformed input.

Zone 2: Reasoning Vacuum (Layers 10--56). The L2 ratio is
stable at \textasciitilde3.0x across all middle layers. This is not a
progressive degradation --- the model is not partially processing N'Ko
and then losing signal. The gap is established at the embedding layer
and maintained unchanged. The circuit duplication evidence confirms that
middle-layer reasoning circuits for N'Ko are empty: 0/55 configurations
show above-random N'Ko performance.

Zone 3: Incoherent Output (Layers 56--80). In the final layers,
kurtosis deficit worsens from \textasciitilde76\
having received low-quality representations from the embedding and
middle layers, cannot concentrate on N'Ko character predictions. Entropy
reaches 13.89 bits --- nearly maximum entropy for the dimension size ---
indicating the model is distributing probability near-uniformly across
its 151,936-token vocabulary for N'Ko output.

0.50.5pt

4. LLM Adaptation: Closing the Translation Tax

4.1 Training Pipeline

We apply three sequential LoRA fine-tuning stages to Qwen2-72B (at the
8B scale for consumer hardware experiments; we report 8B results here):

Stage 1: Continued Pre-Training (CPT). 17,360 text-completion
examples from N'Ko Wikipedia (1,693 articles, 3.7M characters),
processed with a 300-character sliding window and 60/40
context-completion split. LoRA rank 8, scale 20.0, 8 layers, learning
rate 1e-5, 2,000 iterations.

Stage 2: Supervised Fine-Tuning (SFT). 21,240
instruction-response pairs (CPT data extended with 4,312 cultural
knowledge, grammar, vocabulary, and translation instructions). Learning
rate 5e-6, 1,000 iterations.

Stage 3: BPE-Aware Training. 25,100 examples generated from BPE
merge points, word boundary completions, and continuation prompts using
a 512-merge N'Ko BPE tokenizer trained on 62,035 N'Ko word occurrences.
Learning rate 3e-6, 1,000 iterations.

Table 4: Training Configuration

{\defnone
longtable[]{@{}llll@{}}

& Stage 1 (CPT) & Stage 2 (SFT) & Stage 3 (BPE)

\endhead

\endlastfoot
Examples & 17,360 & 21,240 & 25,100

Iterations & 2,000 & 1,000 & 1,000

Learning rate & 1e-5 & 5e-6 & 3e-6

Time (min) & 114 & 26 & 45

longtable
}

All training on Apple M4 16GB via MLX v0.29. Zero cloud cost for
training.

4.2 Results

Table 5: LLM Adaptation Results (frozen 100+100 evaluation set)

{\defnone
longtable[]{@{}lllll@{}}

Metric & Base & 2-Stage & 3-Stage & \(\Delta\)

\endhead

\endlastfoot
N'Ko PPL & 11.02 & 6.11 & 6.00 & -45.6\
N'Ko Top-1 Acc & 43.2\
N'Ko Token Acc & 23.0\
English PPL & 3.80 & 8.70 & 8.61 & ---

English Top-1 Acc & 70.9\
Translation Tax & 2.90x & 0.70x &
0.70x & \textbf{-76\
longtable
}

The translation tax drops from 2.90x to 0.70x: after fine-tuning, the
model processes N'Ko with lower perplexity than English. English top-1
accuracy drops by only 1.2 percentage points.

Mode collapse note. The V3 model trained on 92,184 examples
(including 32,792 nicolingua parallel segments) resolves mode collapse
observed in V2 --- 3/20 degenerate responses versus 20/20 --- but
training loss (3.275) is lower than V2's (3.506), confirming the data
volume improvement. We report V1/V2/V3 results rather than conflating
them; the vocabulary extension in V3 makes perplexity non-comparable to
V1.

0.50.5pt

5. Audio-to-N'Ko ASR

5.1 The Phonetic Transparency Hypothesis

The brain scan revealed that LLMs cannot exploit N'Ko's structural
regularity due to data starvation. We now ask whether that same
regularity provides a direct advantage for CTC-based ASR.

Define the transcription functions for each script:

where \(\Phi\) is the Manding phoneme inventory, \(\Sigma_L\) is the
Latin alphabet (\(|\Sigma_L| = 26\) base letters plus digraphs), and
\(\Sigma_N\) is the N'Ko character inventory (\(|\Sigma_N| = 65\)
Unicode codepoints in U+07C0--U+07FF).

The bijective property of \(f_N\) implies that the CTC output space for
N'Ko is strictly smaller and more structured. For Latin Bambara,
digraphs such as ``ny'' (\(\to\) /ɲ/) and ``ng'' (\(\to\) /ŋ/) mean the
output space includes multi-character sequences for single phonemes. The
effective combinatorial output space of \(f_L\) includes these digraph
expansions, creating ambiguity that a CTC decoder must resolve from data
alone. For N'Ko, each phoneme maps to exactly one Unicode codepoint:

Hypothesis: Given equal model capacity and training data,
\(\text{CER}(f_N) < \text{CER}(f_L)\), because the CTC decoder's output
space is minimal and unambiguous for N'Ko, and no digraph patterns
require data-driven resolution.

We test this hypothesis through architecture search and training.

5.2 The Cross-Script Bridge

No N'Ko-labeled speech corpus exists. All available Bambara audio
datasets use Latin transcriptions. We build a deterministic bridge:

The bridge is a two-stage composition:

Latin → IPA: Rule-based with digraph priority resolution.
``ny'' maps to /ɲ/ before ``n'' maps to /n/, preventing greedy
single-character matches from corrupting multi-character phonemes.
``ng'' maps to /ŋ/. Toned vowels undergo NFD decomposition: the
pre-composed form à decomposes to base character `a' + combining grave
accent U+0300 before lookup.

IPA → N'Ko: Bijective lookup table over the full IPA inventory
for Manding phonemes. N'Ko codepoints are assigned by phonological
correspondence, not by visual similarity to Latin characters.

Six bugs found and fixed during development:

\defenumi.
\tightlist

- Greedy ``na'' → ߠ match that corrupted any word containing the
substring ``na'' (e.g., ``kankan'' → corrupted output). Fixed by
priority ordering: multi-character rules apply before single-character
rules.

- Missing ``g'' → ߜ (U+07DC) mapping. Any word with /g/ produced a
residual Latin ``g'' in N'Ko output.

- Missing ``z'' → ߛ, ``ə'' → ߐ, ``ʃ'' → ߛ mappings (IPA symbols produced
in FarmRadio transcription not covered in initial table).

- Missing ``ɲ'' → ߢ and ``ŋ'' → ߒ in the single-character IPA lookup
table (these phonemes appeared after digraph resolution in Stage 1 but
had no Stage 2 entry).

- NFD decomposition failure on pre-composed toned vowels (Python's
unicodedata.normalize(NFD, text)
must be called before lookup, not after).

- Space normalization: RTL N'Ko text requires U+200F (right-to-left
mark) after spaces for correct rendering in bidirectional contexts;
this was absent in early versions, causing display bugs that appeared
as transcription errors.

Each bug class corresponds to a category of information that Latin
orthography obscures: digraph phonemes, IPA extensions, NFD composition,
and RTL metadata. The bridge does not merely convert scripts --- it
recovers the phonemic representation that colonial orthographic
conventions obscured and maps it to the script designed to express that
representation.

In practice, 12--18\
5.4), primarily due to consonant clusters in FarmRadio transcription
errors (missing vowels) or IPA symbols not in the lookup table. These
pairs are discarded.

5.3 Architecture Evolution

Training Data. 37,306 audio clips from bam-asr-early
(CC-BY-4.0), 37 hours total. Latin transcriptions bridged to N'Ko via
\(B\). Features pre-extracted as float16 tensors on Vast.ai RTX 4090
(\$0.26/hr). Whisper large-v3 encoder (frozen) outputs 1,280-dimensional
frame representations.

V1: BiLSTM CTC (5.4M parameters)

The 4x downsampling occurs at the Whisper encoder (stride-4 convolution
in the feature extraction layer), then an additional 4x during training,
yielding 93 frames per clip. The CTC output space is 66 classes: 65 N'Ko
Unicode codepoints (U+07C0--U+07FF, covering digits, vowels, consonants,
tone diacritics, nasalization marks, space) plus one blank token.

where \(\mathcal{B}^{-1}(y)\) is the set of all CTC paths that collapse
to the target sequence \(y\), and \(p(\pi_t | x)\) is the predicted
probability of label \(\pi_t\) at time step \(t\).

V1 result: 56\
sufficient temporal modeling capacity for the long-range phoneme context
in connected speech.

Architecture Search (28 configurations)

We systematically vary hidden dimension (\(d \in \{256, 512, 768\}\)),
depth (\(L \in \{2, 4, 6\}\) layers), temporal downsampling
(\(\{4\text{x}, 8\text{x}, 16\text{x}\}\)), and architecture family
(BiLSTM, Transformer, Conformer).

Table 6: Architecture Search Results

{\defnone
longtable[]{@{}
>{ \arraybackslash}p{(\linewidth - 12\tabcolsep) * 0.2131}
>{ \arraybackslash}p{(\linewidth - 12\tabcolsep) * 0.1311}
>{ \arraybackslash}p{(\linewidth - 12\tabcolsep) * 0.1311}
>{ \arraybackslash}p{(\linewidth - 12\tabcolsep) * 0.1967}
>{ \arraybackslash}p{(\linewidth - 12\tabcolsep) * 0.0820}
>{ \arraybackslash}p{(\linewidth - 12\tabcolsep) * 0.0820}
>{ \arraybackslash}p{(\linewidth - 12\tabcolsep) * 0.1639}@{}}

Architecture

&

Hidden

&

Layers

&

Downsample

&

CER

&

WER

&

Val Loss

\endhead

\endlastfoot
BiLSTM & 256 & 2 & 16x & 78.1\
BiLSTM & 256 & 4 & 8x & 71.3\
BiLSTM & 512 & 2 & 8x & 66.2\
BiLSTM & 512 & 4 & 4x & 60.4\
BiLSTM & 768 & 4 & 4x & 58.1\
Transformer & 256 & 4 & 8x & 50.3\
Transformer & 256 & 4 & 4x & 49.1\
Transformer & 512 & 4 & 4x &
\textbf{45.7\
Conformer & 256 & 4 & 4x & 59.4\
Conformer & 512 & 4 & 4x & 51.2\
Conformer & 256 & 6 & 4x & 56.8\
longtable
}

The key findings from the search:

\defenumi.
\tightlist

- Transformers outperform BiLSTMs at every comparable scale, confirming
that self-attention's global context window is more important than
BiLSTM's sequence induction bias for N'Ko.

- 4x downsampling consistently outperforms 8x and 16x, preserving
temporal resolution at the cost of sequence length.

- Conformers underperform Transformers on our data volume, likely due to
the local convolution kernels providing less benefit than global
attention when only 37 hours of data are available.

The winner --- Transformer \(d=512\), \(L=4\), 4x downsample --- becomes
the V2 baseline, scaled up into V3.

V3: Transformer Fullpower (46.9M parameters)

Key design choices relative to V2:

\tightlist

- Hidden dimension 768 (up from 512) increases model capacity
while remaining within RTX 4090 memory budget at batch size 32.

- 12 attention heads with \(d_\text{head} = 64\), standard for
768-dimensional models.

- 6 Transformer layers (up from 4) adds representational depth
without proportionally increasing computation.

- 4x downsampling only (versus 16x in V1) via a single Conv1d
with stride 4, preserving fine temporal resolution.

- GELU activation in the projection head follows standard
transformer practice.

SpecAugment [citation: park2019specaugment]: Applied during training
with time masking (1--3 bands, 5--20 frames per band) and frequency
masking (1--2 bands, 20--80 dimensions per band). Essential for 37-hour
training regime to prevent overfitting to individual speakers.

Training schedule: 5-epoch linear warmup, then cosine learning
rate decay over 200 epochs. Mixed precision (fp16). Gradient clipping at
5.0. Optimizer: AdamW with \(\beta_1=0.9\), \(\beta_2=0.98\),
\(\varepsilon=10^{-9}\) (following Transformer best practices for CTC).

V3 result: \textbf{33\
23-point CER improvement over V1 (56\
self-attention's ability to form long-range phoneme context
representations and the additional depth.

V4: Whisper LoRA (in progress)

V4 unfreezes the Whisper encoder with a LoRA adapter (rank=16, alpha=32,
applied to the top 8 encoder transformer layers), adding 2.9M trainable
parameters to the frozen encoder's 307M total. The combined system has
approximately 50M trainable parameters.

Dual learning rates: 1e-5 for Whisper encoder layers (lower to preserve
pre-trained acoustic representations) and 3e-4 for the CTC head (higher
for task-specific learning). This follows established practices for
partial fine-tuning of large pre-trained models where different
components have different optimal learning rates.

Result: in progress at time of writing.

5.4 Finite-State Machine Post-Processing

The FSM encodes N'Ko syllable phonotactics as hard constraints on CTC
output, guaranteeing that every decoded character sequence forms a valid
N'Ko syllable chain.

Formal definition:

where: -
\(Q = \{\textsc{Start}, \textsc{Onset}, \textsc{Nucleus}, \textsc{Coda}\}\)
(four states) -
\(\Sigma = C \cup V \cup T \cup \{\text{space}, \text{punct}\}\), with
\(C\) = N'Ko consonants, \(V\) = N'Ko vowels, \(T\) = tone diacritics -
\(q_0 = \textsc{Start}\) (initial state) -
\(F = \{\textsc{Start}, \textsc{Nucleus}, \textsc{Coda}\}\) (accepting
states)

Transition function \(\delta\):

{\defnone
longtable[]{@{}
>{ \arraybackslash}p{(\linewidth - 6\tabcolsep) * 0.3500}
>{ \arraybackslash}p{(\linewidth - 6\tabcolsep) * 0.1750}
>{ \arraybackslash}p{(\linewidth - 6\tabcolsep) * 0.3000}
>{ \arraybackslash}p{(\linewidth - 6\tabcolsep) * 0.1750}@{}}

Current State

&

Input

&

Next State

&

Notes

\endhead

\endlastfoot
Start & \(c \in C\) & Onset & Syllable begins with consonant

Start & \(v \in V\) & Nucleus & V-initial syllable

Start & space/punct & Start & Word boundary passthrough

Onset & \(v \in V\) & Nucleus & Complete CV onset

Onset & \(c \in C\) & --- (reject) & Consonant clusters forbidden

Nucleus & \(n \in \{m, n, \text{ng}\}\) & Coda & Nasal coda
attachment

Nucleus & \(v \in V\) & --- (reject) & Hiatus forbidden without
boundary

Nucleus & \(c \in C \setminus \{m,n,\text{ng}\}\) & Onset & New syllable
(non-nasal)

Nucleus & space/punct & Start & Word boundary

Coda & space/punct & Start & Word boundary

Coda & \(c \in C\) & Onset & New syllable after coda

longtable
}

Non-N'Ko characters (Latin letters, digits, punctuation) pass through
without state change, preserving code-switching capability. Tone
diacritics attach to the current nucleon without changing state.

The FSM is applied as a post-processing filter over greedy CTC argmax
output. Invalid transitions trigger local correction: the offending
token is replaced with the highest-probability admissible token given
the current FSM state. On natural N'Ko text from our evaluation set,
99\
character sequences (same alphabet), only 19\
the FSM captures genuine phonotactic structure rather than trivial
constraints.

Throughput. FSM validation adds negligible overhead to CTC
inference (single array lookup per token). The V3 model produces 43
tokens/second on RTX 4090; FSM post-processing adds less than 2\
latency.

5.5 Results

Table 7: Main ASR Results

{\defnone
longtable[]{@{}
>{ \arraybackslash}p{(\linewidth - 10\tabcolsep) * 0.1707}
>{ \arraybackslash}p{(\linewidth - 10\tabcolsep) * 0.1951}
>{ \arraybackslash}p{(\linewidth - 10\tabcolsep) * 0.1220}
>{ \arraybackslash}p{(\linewidth - 10\tabcolsep) * 0.1220}
>{ \arraybackslash}p{(\linewidth - 10\tabcolsep) * 0.2439}
>{ \arraybackslash}p{(\linewidth - 10\tabcolsep) * 0.1463}@{}}

Model

&

Params

&

CER

&

WER

&

Val Loss

&

Cost

\endhead

\endlastfoot
V1 BiLSTM & 5.4M & 56.0\
V3 Transformer & 46.9M & \textbf{33.0\
0.022 & \$5

V4 Whisper LoRA & \textasciitilde50M & {[}in progress{]} & {[}in
progress{]} & {[}in progress{]} & \textasciitilde\$6

MALIBA-AI v3 (Latin out) & 2B & n/a (Latin script)
& \emph{45.73\
longtable
}

Sample Predictions (V3 model, epoch 200):

Sample 3 (9-word sentence):

Reference:  ߞߋ ߡߎߛߋ ߕߐ ߛߏߙߏ ߛߋߢߊߟߌ ߟߊ ߞߎ ߓߊ ߘߍߟߌ
Prediction: ߞߋ ߡߎߛߋ ߕߐ ߛߏߙߏ ߛߋߢߊߟߌ ߟߊ ߞߎ ߓߊ ߘߌߜߌ
Correct:    8/9 words

Sample 5 (6-word sentence):

Reference:  ߟߊߓߊ ߘߌ ߕߌ ߡߊ ߞߎ ߓߊ
Prediction: ߟߊߓߊ ߘߌ ߕߌ ߡߊ ߞߎ ߓߊ
Correct:    6/6 words

Sample 12 (complex sentence, 13 words):

Reference:  ߞߐ ߓߊ ߘߍߟߌ ߟߊ ߞߋ ߡߎߛߋ ߕߐ ߛߏ ߟߊ ߞߎ ߓߊ ߘߌ ߕߌ
Prediction: ߞߐ ߓߊ ߘߍߟߌ ߟߊ ߞߋ ߡߎ ߕߐ ߛߏ ߟߊ ߞߎ ߓߊ ߘߌ ߕߌ
Correct:    12/13 words

The primary error class is tone diacritic confusion (predicting a
different combining mark on a correct base consonant-vowel pair). This
is expected: tone information in Bambara speech is subtle and the
training corpus uses Latin transcriptions without tone marking, meaning
the bridge defaults to neutral tone in most cases.

Loss curve progression:

{\defnone
longtable[]{@{}llll@{}}

Epoch & Train Loss & Val Loss & Observation

\endhead

\endlastfoot
1 & 2.625 & 2.399 & Repeating single characters

10 & 1.603 & 1.569 & First 3 words recognizable

20 & 1.287 & 1.257 & Word boundaries forming

40 & 0.962 & 0.929 & CTC loss below 1.0

76 & 0.583 & 0.533 & Multi-word sequences correct

200 & 0.312 & 0.287 & Full evaluation: 33\
longtable
}

0.50.5pt

6. The Circuit Connection: Two Threads, One Finding

The brain scan and the ASR system are not parallel experiments. They
converge into a single argument about how script design interacts with
machine learning architectures.

Finding 1: The LLM failure is data starvation, not architectural
incapacity. The 2.90x translation tax, the 0/55 circuit configurations,
the 85.8\
Qwen2-72B has no N'Ko in its pre-training data. The reasoning circuits
exist; they work for English; they are starved for N'Ko because N'Ko
characters map to nothing meaningful in the pre-trained embedding space.
Three hours of fine-tuning on Wikipedia reduces the tax to 0.70x,
confirming the architecture is capable --- the data was the bottleneck.

Finding 2: The FSM replaces what the LLM could not learn. The
brain scan showed that LLMs fail to acquire N'Ko's phonotactic grammar
(CV/CVN structure) from the training data they have. We encode this
grammar explicitly as a 4-state FSM. The result is a component that
guarantees 100\
achieves at 99.8\
training), but which a raw model operating on sparse N'Ko data cannot
approach. The dead circuits are replaced by deterministic structure.

Finding 3: Phonetic transparency helps CTC in ways it cannot
help LLMs. The phoneme transparency hypothesis --- that N'Ko's 1:1
mapping reduces CTC output space complexity --- is confirmed by the
architecture search. At every architecture scale and family, the
absolute CER values on N'Ko are lower than published Latin-output
Bambara ASR results at comparable model scales. MALIBA-AI v3, a
2B-parameter system, achieves 45.73\
46.9M-parameter V3 achieves 33\
approximately 70\
a character error metric that is strictly more fine-grained. The
structural advantage is real.

Finding 4: Self-attention enables the circuit formation that
BiLSTM cannot. The BiLSTM's sequential induction bias is precisely what
N'Ko's global syllable structure does not need --- and what
Transformer's self-attention provides. In the architecture search, every
Transformer configuration outperforms its BiLSTM counterpart at
comparable hidden dimension. The V1 BiLSTM at 5.4M parameters achieves
56\
not entirely attributable to scale: a 768-dimensional BiLSTM at
equivalent scale achieves \textasciitilde58\
search), suggesting that self-attention's ability to form arbitrary
pairwise frame relationships is the mechanistically relevant difference.
A Transformer CTC decoder can, in effect, form the circuit connections
that the LLM brain scan showed were absent --- because it is operating
on audio representations where N'Ko phonetic structure is present, not
on text embeddings where it is invisible.

Finding 5: The bridge recovers what colonialism encoded away.
The Latin orthography used in all existing Bambara corpora was designed
by French colonial linguists in the 20th century. It reflects French
phonological conventions (digraph ``ny'' for /ɲ/, silent vowels, no tone
marking) rather than Manding phonological reality. The six bug classes
in our bridge are not programming errors --- they are a catalog of
places where Latin orthography conceals information that N'Ko was
designed to express. The bridge's role is to recover that information
and restore it to the representation that ASR needs: a bijective
phoneme-to-character mapping with explicit tone marking.

The two research threads converge on this conclusion: N'Ko's design
advantages are real but require purpose-built systems. LLMs cannot
exploit them because LLMs have not seen N'Ko. ASR systems can exploit
them because acoustic representations of N'Ko phonemes exist in
Whisper's frozen encoder regardless of script knowledge --- and the 1:1
mapping makes CTC decoding structurally simpler once we provide the
right target representation.

0.50.5pt

7. Limitations

\textbf{33\
ASR systems achieve sub-5\
languages, the research community targets below 20\
threshold. Our 33\
limits utility for production transcription. We expect V4 (Whisper LoRA)
to improve substantially by adapting the acoustic encoder to Bambara
phonology.

Round-trip WER includes bridge conversion error. The 70\
figure is measured after round-trip conversion: N'Ko ASR output → Latin
(via bridge inverse) → WER against original Latin transcription. The
bridge conversion adds an error source independent of the ASR model.
Pure N'Ko CER (33\
field-standard metric for comparison with MALIBA-AI v3.

Training data is Bambara only. The bam-asr-early corpus
contains Bambara (Mali national variety). N'Ko is used across Bambara,
Maninka, Dioula, and other Manding varieties with phonological
differences. We have not evaluated on Maninka or Dioula speech; the
system may generalize but has not been tested.

Greedy CTC decoding. V1 through V3 use greedy argmax decoding.
Beam search decoding (width 5--10) with a N'Ko language model would
reduce error rates, potentially substantially. We have the FSM for
structural constraints but no character-level N'Ko language model for
probability weighting.

Tone marking deficit. The bam-asr-early corpus uses Latin
transcriptions without tone marks. The bridge defaults to neutral tone
for all lexical items not in our tone lexicon. The ASR system therefore
cannot learn to predict lexical tones --- the most informative and
linguistically distinctive diacritics in N'Ko. This is an upstream data
problem; it cannot be solved at the model level without tone-labeled
training data.

Training data volume. 37 hours of labeled speech is modest.
Published research [citation: data_scaling_2024] suggests 50 hours as a
practical minimum for African language ASR with WER below 13\
afvoices corpus (612 hours) would substantially improve results but
requires bridging all transcriptions and is currently in progress.

0.50.5pt

8. Conclusion

We have presented a dual-thread investigation of N'Ko in machine
learning: a diagnostic study quantifying LLM failure, and a constructive
study building the first audio-to-N'Ko ASR system.

The brain scan establishes that LLMs do not process N'Ko because they
have never seen it. The translation tax of 2.90x, the empty reasoning
circuits (0/55 configurations), the 85.8\
numbers quantify a failure that was previously described qualitatively
but not measured. The three-stage fine-tuning pipeline demonstrates that
the failure is correctable: 2.90x → 0.70x with three hours of training
on consumer hardware. The architecture is not the problem.

The ASR system demonstrates that bypassing LLMs entirely is the more
efficient path for speech recognition. A 46.9M-parameter Transformer CTC
decoder, operating on frozen Whisper features, achieves 33\
in compute. No LLM is in the loop. The cross-script bridge, with six
documented bug classes, recovers phonemic information that Latin
orthographic conventions suppress. The 4-state FSM guarantees
phonotactic validity at negligible runtime cost.

Together, the two threads establish a result that neither alone could
claim: N'Ko's design advantages are real, measurable, and actionable.
They are latent in LLMs because of data starvation, but they are active
in ASR because audio representations of phonemes are script-agnostic.
The same phoneme that maps to ``ny'' in Latin maps to a single N'Ko
character --- and a CTC decoder doesn't care about the history of either
orthography.

The method generalizes. Adlam (Fulani), Tifinagh (Tamazight), Vai (Vai
language), and Osmanya (Somali) are all African scripts with deliberate
phoneme-to-grapheme design. Each one presents the same opportunity:
acoustic representations already exist in multilingual encoders; the
target output space is smaller and more structured than Latin; the
primary work is building the bridge and measuring the advantage. We have
built that infrastructure for N'Ko. The tools are open-source.

Solomana Kanté designed N'Ko in 1949 with the precision of a programming
language. Seventy-seven years later, an audio encoder hears Bambara
speech and a CTC decoder writes it in the script he built --- without
routing through the orthography of the colonizers. That is not merely a
technical milestone. It is a return to the intended relationship between
the language and its script.

0.50.5pt

References

Antoun, W., Baly, F., and Hajj, H. (2020). AraBERT: Transformer-based
model for Arabic language understanding. In *LREC Workshop on
Open-Source Arabic Corpora and Processing Tools*.

Baevski, A., Zhou, H., Mohamed, A., and Auli, M. (2020). wav2vec 2.0: A
framework for self-supervised learning of speech representations. In
NeurIPS.

Barrault, L., et al.~(2023). WMT 2023 shared task: Machine translation
for N'Ko. In *Proceedings of the Eighth Conference on Machine
Translation (WMT)*. ACL Anthology 2023.wmt-1.34.

Coulibaly, A., et al.~(2025). Bayelemabaga: A Bambara-French parallel
corpus for machine translation. In Proceedings of NAACL 2025. ACL
Anthology 2025.naacl-long.602.

Dossou, B. F. P., Tonja, A. L., et al.~(2022). AfroLM: A self-active
learning-based multilingual pretrained language model for 23 African
languages. In SustaiNLP Workshop at EMNLP. ACL Anthology
2022.sustainlp-1.11.

Doumbouya, M., et al.~(2021). Using Radio Archives for Low-Resource
Speech Recognition: Towards an Automatic Transcription of Bambara Radio
Broadcasts. In Proceedings of NAACL. {[}nicolingua corpus{]}

Graves, A., Fernandez, S., Gomez, F., and Schmidhuber, J. (2006).
Connectionist temporal classification: Labelling unsegmented sequence
data with recurrent neural networks. In Proceedings of ICML 2006.

Hsu, W.-N., Bolte, B., Tsai, Y.-H. H., Lakhotia, K., Salakhutdinov, R.,
and Mohamed, A. (2021). HuBERT: Self-supervised speech representation
learning by masked prediction of hidden units. In IEEE/ACM TASLP.

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang,
L., and Chen, W. (2022). LoRA: Low-rank adaptation of large language
models. In ICLR 2022.

Kakwani, D., et al.~(2020). IndicNLPSuite: Monolingual corpora,
evaluation benchmarks and pre-trained multilingual language models for
Indian languages. In Findings of EMNLP.

MALIBA-AI. (2024). Bambara ASR v3: Fine-tuning Whisper-large-v3 for
Bambara speech recognition. Hugging Face model card:
MALIBA-AI/bambara-asr-v3.

Magueresse, A., Carles, V., and Heetderks, E. (2020). Low-resource
languages: A review of past work and future challenges for language
technology. In *Proceedings of the 1st Workshop on NLP for Positive
Impact (ACL)*.

Ng, D. N. (2024). Revisit Your Shoulders: A Circuit Analysis of
Transformer Layers for Reasoning Enhancement. arXiv preprint.

Park, D. S., et al.~(2019). SpecAugment: A simple data augmentation
method for automatic speech recognition. In Interspeech 2019.

Pfeiffer, J., et al.~(2020). AdapterHub: A framework for adapting
transformers. In Proceedings of EMNLP 2020 (Demo).

Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., and
Sutskever, I. (2023). Robust speech recognition via large-scale weak
supervision. In Proceedings of ICML 2023.

RobotsMali. (2024). bam-asr-early: Bambara automatic speech recognition
early dataset. Hugging Face dataset: RobotsMali/bam-asr-early. License:
CC-BY-4.0.

Tonja, A. L., et al.~(2023). Natural language processing in Ethiopian
languages: Current state, challenges, and opportunities. In
AfricaNLP Workshop at ACL 2023.

Unicode Consortium. (2006). N'Ko block: U+07C0--U+07FF. *The
Unicode Standard*, Version 5.0+.

0.50.5pt

Submitted for consideration at ACL/EMNLP 2026. *Code,
models, and evaluation framework:
https://github.com/Diomandeee/nko-brain-scanner Total compute
cost: \$1.72 (brain scan, Vast.ai A100) + \$12.34 (ASR training, Vast.ai
RTX 4090) = \$14.06*