Ghana recorded 2,949 road traffic deaths and 16,714 traffic injuries in 2025, the highest annual fatality figure in 35 years [1]. Few of these incidents are captured by any health-information system, and existing emergency hotlines and surveillance forms operate in English, which is a first language for approximately 5% of the population [18].

This paper presents VoiceTrace, a five-stage natural language processing (NLP) pipeline that accepts a spoken injury report in any of five Ghanaian languages and returns two outputs from the same call: a spoken first-aid response in the caller's language and a structured, geocoded record for health surveillance. The pipeline operates over a basic phone call and requires no English literacy or smartphone hardware.

Evaluation uses 126 epidemiologically grounded synthetic injury reports across five languages (Twi, Ga, Ewe, Fante, and Dagbani), which together cover the first languages of approximately 85% of Ghana's population [18]. Three evaluation tracks isolate distinct pipeline stages: end-to-end automatic speech recognition (ASR) accuracy (Track 1); translation round-trip fidelity and extraction F1 on clean text (Track 2); and cross-language extraction consistency (Track 3). Outcomes stratify by language. Twi attains macro-F1 = 0.66 with Cohen's κ = 0.80 on clean translated text, and macro-F1 = 0.43 after ASR at a word error rate (WER) of 51.4%. Fante attains macro-F1 = 0.55 with κ = 0.52 on clean text, falling to macro-F1 = 0.34 after ASR (WER 61.0%). Ga, Ewe, and Dagbani record κ < 0.13, mirroring lower BLEU and BERTScore values for those languages and indicating that the binding constraint is the available machine-translation (MT) capacity rather than the extraction architecture.

The system constitutes the first reported end-to-end voice pipeline for injury management and surveillance in any Ghanaian language.