kcode-audio-ingress
kcode-audio-ingress 0.5.3 provides durable automatic audio transcription
with transport-neutral speaker correction. It retains complete caller-owned
WAV bytes, sends overlapping chunks to Gemini for unstructured speaker
analysis, validates a typed GPT parse, scores speaker rows through a
caller-supplied persistent classifier, reconciles a readable transcript, and
durably exposes one complete correction packet.
The embedding application supplies typed audio and text-generation callbacks
and a shared Arc<kcode_speech_classification::SpeechClassifier>. It owns
provider routing, credentials, accounting, classifier storage, object creation,
Kmap storage, and user messaging. This crate adds no provider client, HTTP,
Kmap, Telegram, object-store, or application-framework dependency.
use Arc;
use Utc;
use ;
use SpeechClassifier;
async
The pipeline uses bounded four-minute chunks with 15-second overlap and four
concurrent chunks. Every chunk is sent to gemini-3.1-pro-preview with the
exported GEMINI_SPEAKER_PROMPT_V0_1; raw responses and validated 24-feature
rows are retained. gpt-5.6-sol parses and reconciles the result. A recording
is automatically trained only when every observation passes the strict
confidence/background/runner-up and unique-name gates. Completed status
includes the transcript and correction packet, and confirm_speakers requires
exact observation coverage before applying human-confirmed names.
See Documentation.md for the complete API and Specification.md for persistence and lifecycle guarantees.