How Sume TTS splits a script into sentence ids (. ! ? only)

Sume's script-source API splits each accepted script block at three marks: ., ! and ?. Everything up to and including a mark is one sentence, text after the last mark is a final sentence, and nothing is dropped: the pieces joined together must equal the canonical text, or the request fails with 422 tts_sentence_selection_invalid ("Source segmentation must be lossless.").

Each sentence gets an id sent_ plus a six-digit counter that runs across all blocks of the script (sent_000000, sent_000001, ...). The manifest also gives source_block_index, sentence_index within the block and character_count.

What counts as a boundary

Why this matters for narration

A selection must be whole, contiguous sentences, so a bad cut limits what you can regenerate on its own. If a decimal or abbreviation splits a line, either rewrite it (4 dollars 50) or select both halves together. Whitespace-only pieces attach to the previous sentence of the same block.

Before splitting, text is normalized to Unicode NFC with CRLF and lone CR turned into LF. The Python below reproduces the rule so you can preview ids and lengths offline.

import re, unicodedata

def sentences(blocks):
    out = []
    for b, raw in enumerate(blocks):
        text = unicodedata.normalize("NFC", raw)
        text = text.replace("\r\n", "\n").replace("\r", "\n")
        for piece in re.findall(r"[^.!?]*[.!?]|[^.!?]+\Z", text):
            if not piece.strip():
                if out and out[-1][1] == b:
                    out[-1][2] += piece
            else:
                out.append([f"sent_{len(out):06d}", b, piece])
    return out

for sid, b, t in sentences(["Hi there. It costs $4.50! Ready?"]):
    print(sid, b, len(t), repr(t))

What a cut costs

Segmenting is free. Generating costs $0.0475 per 1,000 characters rounded up per job, so a 14-character sentence alone is still $0.01, the one-cent minimum. Group neighboring sentences into one job when you can.