/**
 * SRT (SubRip) writer.
 *
 * Standard format:
 *   1
 *   00:00:01,000 --> 00:00:03,500
 *   Hello world.
 *
 *   2
 *   00:00:04,000 --> 00:00:06,000
 *   Second cue.
 *
 * Cue numbering is 1-indexed. Timestamps are HH:MM:SS,mmm with comma decimal.
 * CRLF line endings are NOT required by the spec — most parsers accept LF
 * (we emit LF for simplicity and cross-platform sanity).
 *
 * ── Precision policy ─────────────────────────────────────────
 * SRT's wire precision is exactly 1ms. We canonicalise every timestamp to a
 * non-negative INTEGER millisecond at the boundary of this module:
 *
 *   secondsToMs(s) = Math.max(0, Math.round(s * 1000))
 *
 * All downstream math (cue ordering, gap extension, end>start validation,
 * grouping) runs on those ms-ints. This eliminates two whole classes of
 * float-arithmetic bugs:
 *
 *   1. Cues whose float `end > start` but whose ms-rounded values collide
 *      (e.g. start=0.1001, end=0.1004 → both round to 100). The float check
 *      would pass, then we'd emit `00:00:00,100 --> 00:00:00,100` — invalid.
 *   2. Drift when extending a cue's end up to the next cue's start: float
 *      assignment can carry a 1e-15 epsilon that flips a Math.round at the
 *      ms boundary. Integer assignment is exact.
 *
 * Callers pass seconds (the natural unit for whisper / OpenAI output); the
 * conversion happens exactly once per timestamp, here. If you already have
 * ms-ints upstream, use `formatSrtTimeMs` directly to skip the round trip.
 */
export interface SrtCue {
    /** Start time in seconds. Internally rounded to ms-int. */
    start: number;
    /** End time in seconds. Must be > start AFTER ms-rounding (see policy above). */
    end: number;
    /** Caption text. Multi-line OK. Trimmed. */
    text: string;
}
/**
 * A word with its own timing. Used by `buildWordLevelSrt` to emit one cue per
 * word (TikTok / Reels burned-caption style).
 */
export interface WordCue {
    start: number;
    end: number;
    text: string;
}
/**
 * Word-level SRT — one cue per word. Useful for vertical-format burned-in
 * captions where each word pops as it's spoken.
 *
 * Behaviour:
 *   - Skips empty / zero-duration words (post-rounding, so 0.1001s/0.1004s
 *     pairs that collapse to the same ms are dropped instead of producing
 *     invalid same-stamp cues).
 *   - Optional `groupSize` clusters N consecutive words into one cue (e.g. 2-3
 *     words at a time for a more readable rhythm). Default 1.
 *   - Optional `gapSec` extends a cue's end up to the next cue's start so the
 *     caption stays on screen until the next word appears (closed-caption look).
 */
export declare function buildWordLevelSrt(words: WordCue[], opts?: {
    groupSize?: number;
    gapSec?: number;
}): string;
/**
 * Build an SRT string from cues. Cues are emitted in input order. Throws on
 * cues that collapse to zero or negative duration AFTER ms-rounding —
 * upstream code is expected to handle that ahead of time.
 */
export declare function buildSrt(cues: SrtCue[]): string;
/** Format seconds → "HH:MM:SS,mmm" (SRT timestamp). Single rounding step. */
export declare function formatSrtTime(seconds: number): string;
/**
 * Format ms-int → "HH:MM:SS,mmm". Use this when you already have ms-int
 * timing (e.g. straight off whisper.cpp's `offsets.from`) to avoid an
 * unnecessary round-trip through float seconds.
 */
export declare function msToSrtTime(ms: number): string;
/**
 * Convert float seconds → integer milliseconds with the same rounding rule
 * used everywhere in this module. Negative inputs and NaN clamp to 0.
 */
export declare function secondsToMs(seconds: number): number;
//# sourceMappingURL=srt.d.ts.map