A production voice chatbot pipeline was built for Hausa, Yoruba, Igbo, and English using five independent microservices, each exposing an OpenAI-compatible HTTP endpoint. For speech-to-text, the base whisper-small-multilingual checkpoint produces 30–40% WER on Hausa without fine-tuning; after fine-tuning on approximately 50 hours of native speaker audio per language, WER drops below 12%. In production, 15–25% of incoming audio is background noise or dead air, causing Whisper to hallucinate. The fix uses Silero VAD before the model: if no speech is detected, the request returns an empty transcription immediately. For language identity, Whisper's decoder is hard-forced to the correct language token for English, Hausa, and Yoruba using forced_decoder_ids; Igbo is intentionally excluded because it is not in Whisper's vocabulary. Transcription is streamed as Server-Sent Events, emitting each 30-second chunk as it completes. For translation, Meta's NLLB-200 is used rather than an LLM. The unquantized PyTorch NLLB model runs at ~6 tokens/second on CPU; converting it to CTranslate2 INT8 format yields a converted model that runs at ~150ms.