02Architecture
Six-layer architecture, three original modules
Audio leaves the cellular module digitally and is processed on the customer's edge computer. Nothing goes to the cloud.
- L1
Audio & Call Layer
analog → digitalSIM7600G cellular module answers the call; audio is passed to the edge computer digitally over PCM/I2S. Analog noise is eliminated at source.
SIM7600G
- L2
Speech Recognition Layer
8 kHz → 16 kHz8 kHz phone band is adapted for the 16 kHz model; a quantised Whisper / wav2vec 2.0 model runs locally.
Whisper-tiny
- L3
Turkish Proper-Name Correction Layer
biasing · phoneticsThree stages: contextual biasing → Turkish-specific phonetic matching → dialog confirmation (confidence threshold 0.5).
confidence ≥ 0.5
- L4
Conversation & Workflow Engine
YAML rulesRequired fields, question order and business rules live in YAML configuration; new sectors onboard without code changes.
state machine
- L5
Record Integrity Layer
schema · enum · sourceModel output must conform to a predefined schema; enumerated fields are constrained; every persisted value must literally appear in the transcript; records cannot open with missing mandatory fields.
pre-commit
- L6
Speech Synthesis Layer
Piper TRPiper Turkish voice model runs locally; XTTS-v2 is evaluated as a quality reference.
on-device
output · structured record (YAML schema)
Original modules
These three components are not present in off-the-shelf models; they fill gaps left by the literature.
Turkish-specific phonetic encoding
Phonetic coding algorithms exist for English (Soundex, Metaphone), German (Kölner Phonetik), Slavic (Daitmchotoff) and New Zealand English (Caverphone); no established counterpart for Turkish. The project fills that gap.
Free-form location → administrative unit resolver
The ambiguity created by identical neighbourhood names across provinces is resolved through in-dialog confirmation, without asking the caller administrative questions.
Pre-record integrity guard
Generative-model output is gated by schema constraints, enumerated value sets and source verification, preventing fabricated records.