Korean voiceover: 1,200 Hangul characters cost 5.7 cents, so use NFC

Sume TTS is priced at $0.0475 per 1,000 transcript characters, with spaces and punctuation counted, so a Korean script of 1,200 characters costs 1.2 x $0.0475 = $0.057, or 5.7 cents. The catch is that the same-looking Korean text can contain a very different number of characters. Hangul can be stored as one precomposed syllable or as two or three separate jamo, and text copied from some sources arrives in the decomposed form.

Per-character pricing and the 20,000-character request limit are from Sume's public catalog, read on 2026-10-09. How Sume counts decomposed text is not documented, so normalize before you submit.

Why the count can differ

Unicode has a composed form (NFC) in which one syllable is a single code point, and a decomposed form (NFD) in which it is split into its parts. A syllable with a final consonant is three jamo in NFD. If a counter sees code points, an NFD script of 1,200 visible syllables could count up to 3,600 characters, and the price would triple.

A pre-flight check

Normalize the text to NFC in your code before you count it or submit it, and count after that step. In Python, unicodedata.normalize with the form NFC does it. Then compare len of the text with your budget, and keep it under 20,000.

Also set the language. Sume's guidance for Korean is to send a Hangul transcript with language set to ko; a Korean transcript without a language is mispronounced, and a voice-language mismatch returns a warning that you must confirm before the job is accepted. Never translate a Korean request into English text just to save characters.

import unicodedata

def tts_chars(text):
    clean = unicodedata.normalize("NFC", text)
    return clean, len(clean)

nfd = unicodedata.normalize("NFD", "안녕하세요 여러분")
clean, n = tts_chars(nfd)
print(len(nfd), n, round(n / 1000 * 0.0475, 5))

Budget examples

A 600-character caption script is $0.0285. A 20,000-character script, the largest single request, is $0.95. These numbers assume NFC input and the catalog rate; the TTS tool in Sume's hosted MCP server has a dry_run option that previews cost, so test a short script first and compare the preview with your own count.

• Normalize to NFC, then count.
• Send language ko with a Hangul transcript.
• Stay under 20,000 characters per request.

Where decomposed text comes from

Decomposed Hangul shows up when text is copied from certain file systems, some older documents, and some editors that store filenames or form input in the decomposed form. It looks identical on screen, so you cannot see the problem by eye. Only a character count or a normalization call reveals it.

The fix costs nothing. Add the normalization step to the same function that strips stray whitespace and checks the 20,000-character limit, and every script passes through it. Then the cost you estimate from the Hangul you see is the cost you pay.