kcode-audio-ingress

kcode-audio-ingress 0.4.0 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 persistent classifier, reconciles a readable transcript, and durably exposes one complete correction packet.

The embedding application supplies typed audio and text-generation callbacks. It owns provider routing, credentials, accounting, object creation, Kmap storage, and user messaging. This crate adds no provider client, HTTP, Kmap, Telegram, object-store, or application-framework dependency.

use chrono::Utc;
use kcode_audio_ingress::{
    AudioIngress, AudioInput, AudioTranscriber, ObservationConfirmation,
    RecordingConfirmation,
};

async fn accept(
    transcriber: AudioTranscriber,
    wav_bytes: Vec<u8>,
) -> Result<(), kcode_audio_ingress::Error> {
    let ingress = AudioIngress::open("./audio-ingress", transcriber).await?;
    let submission = ingress
        .submit(AudioInput {
            user_id: "user-root-id".into(),
            bytes: wav_bytes,
            recorded_at: Utc::now(),
            original_filename: Some("recording.wav".into()),
        })
        .await?;

    if let Some(packet) = ingress
        .status()?
        .recordings
        .into_iter()
        .find(|recording| recording.id == submission.recording_id)
        .and_then(|recording| recording.correction_packet)
    {
        let confirmations = packet
            .chunks
            .iter()
            .flat_map(|chunk| &chunk.observations)
            .map(|observation| ObservationConfirmation {
                observation_key: observation.observation_key.clone(),
                confirmed_full_name: "Confirmed Full Name".into(),
            })
            .collect();

        ingress.confirm_speakers(RecordingConfirmation {
            recording_id: packet.recording_id,
            observations: confirmations,
        })?;
    }

    Ok(())
}

Pipeline

Chunking remains bounded to four-minute windows, 15-second overlap, and four concurrent chunks.
Each chunk goes directly to gemini-3.1-pro-preview with the exact exported GEMINI_SPEAKER_PROMPT_V0_1; AudioChunkRequest::schema is None.
The complete raw Gemini response is retained verbatim. A separate gpt-5.6-sol xhigh text callback converts it to validated machine JSON containing utterances, notes, clip validity, ISO 639-3 language, and exactly one typed 24-feature row per local speaker.
The persistent classifier is opened beneath the ingress root. Read-only scoring uses a nonaccepting finite threshold and explicit deterministic probe keys. Clean recordings alone train their deterministic correction keys.
Reconciliation receives explicit chunk-local mappings, marks unclean chunks uncertain, and is forbidden from guessing real identities.
Completed status includes the transcript and complete transport-neutral correction packet. confirm_speakers validates exact observation coverage, updates classifier assignments idempotently, and persists confirmed names.

Version 0.4.0 intentionally breaks the former structured-audio callback: AudioChunkRequest::schema changed from serde_json::Value to Option<serde_json::Value>. See Documentation.md for the complete API and Specification.md for behavioral and persistence guarantees.

kcode-audio-ingress 0.4.0

kcode-audio-ingress

Pipeline