kalandengit/offline_nko_voice_transcript

API-free offline Bambara voice transcription and N'Ko learning tools # offline_nko_voice_transcript API-free Bambara/Manding voice transcription and N'Ko tools. Runtime inference uses only a local MMS Bambara model directory (`local_files_only=True`), the bundled French→N'Ko lexicon, and deterministic Latin→N'Ko transliteration. ```bash pip install -e . offline-nko text "i ni ce" offline-nko search maison pip install -e '.[asr]' offline-nko transcribe sample.wav --model models/mms-bam ``` The model weights are not bundled because they are several gigabytes. Copy a licensed MMS Bambara model into `models/mms-bam` once; inference never calls an API or downloads missing files. The retrieval/update module ranks exact, prefix, then substring matches. `offline-nko update-lexicon reviewed.json updated.json` creates an auditable offline pack from reviewed French-based additions. Fine-tuning should remain a workstation process using licensed, native-reviewed Bambara audio/transcripts. Design inspiration: FluidVoice's language-first onboarding, visible local-model state, live feedback, optional history, and explicit privacy boundaries. This repository contains independent code and does not copy GPLv3 FluidVoice source.