/**
 * Split off the leading run of complete sentences from `text`, returning the
 * speakable prefix and the still-incomplete remainder. A segment is "complete"
 * when it ends in terminal punctuation (optionally followed by closing quotes).
 */
export declare function takeCompleteVoiceText(text: string): {
    text: string;
    remaining: string;
};
export declare function isCompleteVoiceText(text: string): boolean;
/** Join two voice-text fragments, normalizing whitespace at the seam. */
export declare function appendVoiceText(existing: string, next: string): string;
/**
 * Normalize a speakable segment so TTS does not read formatting aloud.
 *
 * LLM output is full of markdown — `**bold**`, `## heading`, `` `code` ``, `[text](url)` —
 * and a TTS engine narrates the punctuation literally ("star star bold star star"). This is
 * the single highest-frequency voice bug class and has no app-layer fix, because the text is
 * generated inside the pipeline. Vapi ships a 14-step "voice formatting plan" on by default;
 * this is the locale-free core of it: markdown removal.
 *
 * Deliberately conservative — it strips *formatting* markers, never content, so a sentence
 * that legitimately contains an asterisk or a hash in prose is left intact where ambiguous.
 * Number/currency/date verbalization is locale-sensitive and intentionally NOT done here yet;
 * doing it wrong (wrong locale, wrong magnitude) speaks worse than leaving the digits. That is
 * a separate stage, not a silent omission.
 */
/**
 * Strip leaked tool-call protocol tokens so TTS never speaks them.
 *
 * A model emits tool calls in its own syntax (`<|tool_call|>…`, `<tool_call>…</tool_call>`,
 * `[TOOL_CALLS]`, harmony `<|channel|>commentary`). The inference server is supposed to parse
 * that into the structured `tool_calls` field; when the serving stack lacks the right parser
 * the tokens fall through as ordinary assistant text and the pipeline speaks the markup aloud.
 * LiveKit's finding: the same weights score 100% behind one endpoint and 0% behind another —
 * it is the serving stack, not the model. Syrinx (and the Kuralle runtime) both read only the
 * finalized structured tool-call and never text-scrape, so neither catches a leaked one.
 *
 * This is a high-precision guard: it removes only unambiguous sentinel tokens/blocks that never
 * occur in real speech. It does NOT try to strip bare JSON function calls — that is ambiguous and
 * stripping legitimate content is worse than the rare leak. A leak of that shape is a serving-stack
 * bug to fix at the endpoint (LiveKit's "one curl" diagnosis), not something to paper over here.
 */
export declare function stripLeakedToolCalls(text: string): string;
export declare function normalizeForSpeech(text: string): string;
/**
 * The FIRST spoken fragment of a turn, taken earlier than a full sentence.
 *
 * Only the first dispatch gates time-to-first-audio: every later sentence is
 * synthesised while earlier audio is already playing, so its buffering is free.
 * But the sentence rule charges the first chunk the full wait for a terminator —
 * measured at ~200ms of a ~950ms TTFA.
 *
 * This takes a leading fragment at a CLAUSE boundary (comma, semicolon, colon,
 * dash) once it is long enough to be worth speaking. Never mid-word: the split is
 * always at punctuation followed by whitespace.
 *
 * Deliberately conservative about the cases `isCompleteVoiceText` guards — it will
 * not split on a decimal point or an abbreviation dot, because those are not
 * clause boundaries and never match here.
 *
 * Returns empty when no safe early split exists, in which case the caller keeps
 * waiting for a full sentence.
 */
export declare function takeFirstFragment(text: string, minChars: number): {
    text: string;
    remaining: string;
};
//# sourceMappingURL=voice-text.d.ts.map