Foreign names in an English voiceover: respell or split the job

For a few foreign names inside an English narration, keep one English job: send language: "en", an English voice, and spell each name the way it should sound. Sume TTS takes one language and one voice per request, and has no inline language tag, so the engine reads every word with that language's rules. Respelling "Siobhan" as "shi-VAWN" adds a few characters, a fraction of a cent at $0.0475 per 1,000. If a name must sound fully native, split the script into jobs and join the audio with Timeline audio, at $0.01 flat per join.

The request fields and the language rule come from the Sume API reference; the join from Timeline audio, the price from API pricing and the polling steps from Jobs and results, read on 2026-10-03. The voice check is read from Sume's code. Whole sentences in a second language are a different job, covered in mixed-language scripts; this page is about names and short phrases inside one language.

What does Sume do with a foreign name in the script?

Nothing special. In current code, the 409 tts_voice_language_mismatch check compares the voice's primary language with the request's language. It never reads the transcript for foreign words, so an English voice with en and a Spanish name raises no error: the name is read by English rules, and only your ear tells you if it landed.

One trap: when language is omitted, Sume infers ko or ja if Hangul or kana letters outnumber Latin letters. A short English line carrying a long Korean name can flip, then meet an English voice and stop with a 409. Always send language.

Which approach fits which name?

Respelling keeps one voice, one job and one flat read. Splitting changes the voice at each seam, because a voice's language is part of the check, so pick it only when respelling cannot get there.

• The schema also lists an optional pronunciation_dict_id; the pages read do not say how to create one, so Text to speech pronunciation is the place to look.
• In current code Sume's TTS wav output is written as one channel, so two TTS wav parts join cleanly; a mismatch fails as audio_parts_channel_mismatch.

How do I respell names in a script?

Keep a map of names to spellings, apply it before you send, and check the language guard offline. This script runs without a key and prints the request body, the character count and the cost.

import json, re
RESPELL = {"Joaquín": "wah-KEEN", "Siobhan": "shi-VAWN", "Niamh": "neev"}
script = "Joaquín and Siobhan met Niamh at noon."
def respell(text):
    for name, say in RESPELL.items():
        text = re.sub(rf"\b{re.escape(name)}\b", say, text)
    return text
def inferred(text):
    hangul = len(re.findall("[가-힣]", text))
    kana = len(re.findall("[぀-ヿ]", text))
    latin = len(re.findall("[A-Za-z]", text))
    return "ko" if hangul > latin else "ja" if kana > latin else None
spoken = respell(script)
assert inferred(spoken) is None  # with language omitted this stays English
body = {"transcript": spoken, "language": "en",
        "voice": {"id": "voi_0123456789abcdef0123456789abcdef"}}
print(json.dumps(body, ensure_ascii=False))
print(len(spoken), "characters =", round(len(spoken) * 0.0475 / 1000, 5), "USD")

How do I check the result?

Generate only the sentence with the name, listen, and adjust the spelling. Every hyphen and capital counts as a character, but a retake of one line is cheap. Check a TTS take with an STT round trip catches dropped words, though not a mispronounced one.