Retrieval-Centric ASR for N'Ko: Exploiting Script Structure to Beat Sequence-to-Sequence

We present a retrieval-centric automatic speech recognition (ASR) architecture for Bambara, targeting N'Ko script output directly rather than routing through Latin transcription. The central insight is structural: N'Ko enforces a strict 1:1 phoneme-to-grapheme mapping, explicit tonal diacritics, and a mathematically complete syllable inventory of 3,024 entries (all V, VN, CV, and CVN patterns across five tones). This finite, well-structured output space makes retrieval a better fit than sequence-to-sequence decoding. Our pipeline freezes a Whisper encoder to extract audio embeddings, projects them into a shared 512-dimensional space alongside N'Ko syllable embeddings, and retrieves the nearest codebook entry at each step. A 4-state finite-state machine (FSM) encoding N'Ko phonotactics constrains beam search during assembly, guaranteeing that every output token sequence forms a valid N'Ko syllable chain. Training data comes from two YouTube sources: 1,461 episodes of Djoko dialogue (audio + FarmRadio Whisper transcription, bridged to N'Ko) and 532 babamamadidiane teaching videos (dynamic scene detection + Gemini 3 Flash OCR for on-screen N'Ko extraction). The current best published result for Bambara ASR is MALIBA-AI bambara-asr-v3 at 45.73% WER on Latin-script output. Our architecture bypasses Latin entirely. Quantitative results on real audio will be reported as training completes. **Keywords:** low-resource ASR, N'Ko, Bambara, Manding, retrieval-augmented, finite-state machine, CTC, script-structure exploitation