On-Device Voice Command Recognizer

We're building a voice-first learning app for non-literate users, demoed in mid-July. We need an ML/speech engineer to build a small, reliable, offline on-device voice command recognizer for three low-resource African languages: Mooré, Dioula, Fulfulde. This is a closed-grammar problem (~30 fixed commands), not open transcription — command-spotting / keyword detection, not a chatbot. What you'll build: Audio preprocessing pipeline + a documented data-collection protocol (recording is run locally by native-speaker coordinators; you specify the spec and process the data). Fine-tuned compact command models (e.g. MMS / wav2vec2 / W2v-BERT-2.0 or a keyword-spotting model), per language. Phased depth: Mooré polished; Dioula + Fulfulde at functional bootstrap. On-device export (ONNX / TFLite) running on Android fully offline — no cloud APIs. Evaluation report (per-command, per-language) + a documented improvement plan. Scope boundary: you deliver the recognizer + on-device export through a clean local inference interface (audio + active command set + language → command_id, confidence, accept/confirm/reject). Dialogue logic, UI and audio prompts are handled by our team. (A separate praise-audio playback feature is out of scope — no ML.) Acceptance test: on a target Android phone in airplane mode, the sample recognizes the 30 commands in the three languages at the reported accuracy and latency. Engagement: fixed USD 4,000, milestone-based (3 milestones), start ASAP, ~4–6 weeks, hands-on collaboration with the founder. A short paid test task (1–2 h) may precede commitment. To apply, briefly tell us: (1) a relevant speech / low-resource / on-device project you've shipped, and (2) your approach to a closed-grammar, offline, multi-language recognizer. A full Statement of Work will be shared with shortlisted candidates.