/**
 * Second-opinion re-transcription for DIRECT koi addresses.
 *
 * In a noisy room the realtime STT reliably catches the wake word (context
 * biasing) but garbles the QUESTION — 2026-07-07 the owner asked "Koi, where
 * does that water live?" and the live stream delivered "Koi, what is it?
 * What are the." The koi then confidently answers a question that was never
 * asked. The SAME audio re-run through Whisper recovered the real question,
 * so: when the owner addresses the koi, re-transcribe just that utterance's
 * audio through Groq Whisper (concurrently with the context load — no added
 * latency) and hand BOTH hearings to the responder.
 *
 * Audio sources, in order:
 *   1. the per-utterance PCM clip (precise, but the voice-ID bridge deletes
 *      clips seconds after embedding — often already gone by eval time),
 *   2. the 60s rolling WAV buffer chunk covering the utterance timestamp
 *      (lives 24h; upload the chunk, keep only segments overlapping the
 *      utterance window).
 */
/**
 * Preserve the DIARIZED per-utterance clip the moment its trigger arrives.
 * The voice-ID bridge reaps clips seconds after embedding, so by eval time
 * the precise single-speaker audio is usually gone and the second opinion
 * would fall back to UNSEPARATED room audio — losing exactly what Soniox's
 * diarization+endpointing provides. Copy is a few hundred KB; stale
 * snapshots are pruned on each call. Returns the snapshot path, or null.
 */
export declare function snapshotClip(audioPath: string | undefined): string | null;
/** Wrap raw 16 kHz mono s16le PCM (the daemon's clip format) in a WAV header. */
export declare function wavFromPcm16(pcm: Buffer, sampleRate?: number): Buffer;
/** Find the rolling-buffer chunk whose 60s window covers `ts`. */
export declare function findBufferChunk(ts: number, dir?: string): {
    path: string;
    startMs: number;
} | null;
export interface SecondOpinion {
    text: string;
    /** "clip" = the diarized single-utterance audio (trustworthy attribution);
     *  "window" = unseparated room audio around the moment (may mix speakers). */
    source: "clip" | "window";
}
/**
 * Re-transcribe the addressed utterance's audio and return the alternate
 * hearing when it MATERIALLY differs from the realtime transcript. Returns
 * null when the second engine agrees, the audio is gone, or Groq is
 * unavailable — callers treat null as "nothing to add".
 */
export declare function secondOpinionTranscript(opts: {
    text: string;
    ts: number;
    audioPath?: string;
    bufferDir?: string;
    apiKey?: string;
    timeoutMs?: number;
    fetchImpl?: typeof fetch;
}): Promise<SecondOpinion | null>;