/**
 * Media-echo classification for mic transcripts.
 *
 * When device audio plays over the speakers (a video, a call, music), the mic
 * hears it too and Soniox happily transcribes it — so media dialogue shows up
 * as mic speech attributed to the wearer, and wake-words spoken IN the media
 * ("koi") fire the proactive pipeline. This module classifies a mic_transcript
 * segment as MEDIA ECHO when its text substantially matches recent system
 * audio_transcript (ScreenCaptureKit device-audio) text.
 *
 * Exact matching cannot work: Soniox segments the two streams on different
 * boundaries and decodes them independently, so the same sentence arrives as
 * differently-shaped segments with small word-level diffs. Instead we use
 * token containment: a rolling window of recent system-audio tokens, and a
 * mic segment is echo when enough of its significant tokens appear in that
 * window. If the system stream is silent (headphones, nothing playing) the
 * window is empty and nothing is ever classified as echo.
 *
 * Used by BOTH capture paths — context-block rendering (PATH A) and proactive
 * triggers (PATH B). Each consumer owns its own tracker instance because each
 * tails the audio JSONL at its own offset/cadence.
 */
/**
 * How long system-audio tokens stay matchable. Soniox can finalize a long mic
 * utterance tens of seconds after the media line it echoes (the two streams
 * flush on different boundaries), so the window must comfortably cover one
 * long utterance either way — 60s.
 */
export declare const SYSTEM_AUDIO_WINDOW_MS = 60000;
/**
 * Mic segments with fewer significant tokens than this are never classified
 * as echo — one or two shared words ("yeah", "okay cool") is not evidence,
 * and misclassifying real user speech is worse than letting an echo through.
 */
export declare const MIN_SIGNIFICANT_TOKENS = 3;
/**
 * Containment ratio at/above which a mic segment counts as echo. ~70%: the
 * two streams' independent decodes disagree on a few words per sentence, so
 * requiring 100% would never fire; much lower and ordinary topical overlap
 * (the user talking ABOUT the video they're watching) starts false-positiving.
 */
export declare const ECHO_CONTAINMENT_THRESHOLD = 0.7;
/** Lowercase, strip punctuation, split into word tokens. */
export declare function normalizeTokens(text: string): string[];
/** The tokens that count toward echo evidence (see MIN_TOKEN_LENGTH). */
export declare function significantTokens(text: string): string[];
/**
 * Fraction of `tokens` (deduped) present in `window`. Deduping keeps one
 * repeated common word from dominating the ratio.
 */
export declare function containmentRatio(tokens: string[], window: ReadonlySet<string>): number;
/**
 * Stateful rolling window of recent system-audio tokens + echo classification
 * for mic segments. Feed every system audio_transcript line via
 * `noteSystemAudio`; ask `isEcho` for each mic_transcript line. Within one
 * batch of JSONL lines, feed ALL system lines before classifying mic lines —
 * a mic echo can land in the file before its system-audio source segment.
 */
export declare class EchoTracker {
    private entries;
    noteSystemAudio(text: string, ts?: number): void;
    /** True when the mic segment substantially matches recent system audio. */
    isEcho(text: string, ts?: number): boolean;
    private prune;
}
/**
 * How long a mic segment votes for speaker dominance. Long enough that the
 * wearer stays "you" through someone else's interjections; short enough that
 * handing the device to someone else re-baselines within minutes.
 */
export declare const SPEAKER_DOMINANCE_WINDOW_MS: number;
/**
 * Rolling dominant-speaker vote over NON-echo mic segments. The wearer talks
 * into their own mic far more than anyone else, so the most frequent Soniox
 * `speaker` tag over the recent window is "you"; every other tag is another
 * voice in the room.
 */
export declare class DominantSpeakerTracker {
    private entries;
    note(speaker: string, ts?: number): void;
    /** The most frequent recent speaker tag, or null when nothing is tallied. */
    dominant(now?: number): string | null;
    private prune;
}