How Fahan Was Built

Before any audio processing begins, Fahan validates the device, checks subscription status, and retrieves or creates a conversation session. The session remembers which languages are active, what direction translation is flowing, and whether real speech has been detected. This isn't security theater — it's a stateful conversation manager that prevents ghost sessions and ensures every turn knows its context.

Raw audio is fed into a neural speech recognition system specifically tuned for Somali phonetics. This isn't the same engine used for English — Somali has unique vowel harmonies, retroflex consonants, and tonal patterns that generic models butcher. The transcription engine also handles code-switching, where a Somali speaker drops English words mid-sentence — something that happens constantly in diaspora speech and that most systems choke on.

This is where Fahan diverges from every competitor. Instead of assuming a fixed translation direction, Fahan detects WHICH language is being spoken on every single utterance and dynamically routes the translation. If a Somali speaker talks, it translates to English. If the English speaker responds, it flips. This dual-rail system means two people can have a natural back-and-forth conversation without ever touching a button or switching a mode.

The transcribed text is translated by a model selected specifically for the detected language pair. But translation alone isn't enough — Fahan runs every output through a quality validation pipeline. Empty results are caught. Repetitive garbage patterns are detected and rejected. A quality score is calculated for every translation. If a translation fails validation, the system gracefully falls back to echoing the original text rather than serving garbage to the user.

The translated text needs to be spoken aloud — but not by a generic robot voice. Fahan routes to different voice synthesis engines depending on the target language. Somali output goes to a specialized engine trained on natural Somali speech cadences. All other languages route to a separate high-fidelity engine. The result sounds like a real person speaking, not a machine reading text.

Every successful, quality-validated translation is cached using a composite key that encodes the translation direction, the voice engine used, and a normalized version of the input text. If the same phrase is spoken again — even by a different user — the cached audio is served instantly, cutting latency to near zero. The cache also strips background noise labels, normalizes whitespace, and deduplicates requests to prevent double-processing from network retries.