Per-request character limits: Eleven v4 10,000 vs Sume TTS 20,000

Eleven v4 accepts a single generation of up to 10,000 characters, and the Sume TTS surfaces accept a transcript of 1 to 20,000 characters. For a 45,000-character script that means at least 5 requests on Eleven and 3 on Sume (ceil of 45,000 / 10,000 and of 45,000 / 20,000). Fewer requests means fewer seams to hide, but the limit is only one part of keeping a long read consistent.

The limits side by side

Splitting a script on sentence boundaries

Never cut mid-sentence: the seam lands in the middle of a phrase and the intonation resets. This script splits on sentence ends and keeps each chunk under a limit you choose. It splits a single oversized sentence hard, as a last resort.

import asyncio
import re


def chunk(text, limit):
    out, cur = [], ''
    for s in re.split(r'(?<=[.!?])\s+', text.strip()):
        while len(s) > limit:
            if cur:
                out.append(cur)
                cur = ''
            out.append(s[:limit])
            s = s[limit:]
        if cur and len(cur) + len(s) + 1 > limit:
            out.append(cur)
            cur = s
        else:
            cur = (cur + ' ' + s).strip()
    return out + ([cur] if cur else [])


async def main():
    script = 'This is one sentence. ' * 2250  # about 49,500 characters
    for limit in (10000, 20000):
        parts = chunk(script, limit)
        print(limit, len(parts), max(len(x) for x in parts))


asyncio.run(main())

Source binding in Sume

For scripted work Sume lets you send transcript_source instead of raw text: a script_revision_id plus ordered sentence_ids from GET /v1/tts-1.0/source. The API resolves the IDs against the accepted revision and rejects duplicates, reordered or unknown IDs. A source-bound job returns a receipt, and the reference states plainly that a receipt proves the submitted text, not the pronunciation. You still listen to the result. Note that GET /v1/tts-1.0/source works on an authenticated thread; for plain scripts you send transcript and chunk it yourself.

Whichever engine you use, render the seams. Play the last two seconds of one chunk into the first two of the next and check pace and level before you approve a long read.