Text Normalization

One flag, three positions — and a billing guarantee worth knowing: normalization changes what is SPOKEN, not what is billed or recorded. You are charged on the characters you sent (verbalising "$1,234.50" into fifty-odd characters of words does not turn a 9-character request into a 55-character charge), and moderation, history and the idempotency fingerprint all see your original text. with_timestamps aligns against what was spoken, so the timings line up with the verbalised words. SSML is the surgical alternative (second tab): set "ssml": true and mark up spans by hand — <say-as> for characters/digits/ordinal/date/telephone readings, <sub alias="…"> to substitute, <break time="750ms"/> for real timed silence (the bracket forms [pause], [pause:750ms], [pause:1.5s] convert to breaks too, clamped 50 ms–10 s), plus <prosody> and <emphasis>. <phoneme> is parsed but NOT applied — no deployed engine consumes phonetic input, so the grapheme text is spoken; use a respelling instead. SSML is buffered-only: the streaming endpoint takes plain text.

Scope

tts:synthesize

Credits

⌈characters sent / 100⌉ × 100 — the verbalised expansion is never billed

# "on": the price is read as money, the date as a date.
curl https://staging.sonicvox.ai/api/v1/text-to-speech \
  -H "sv-api-key: $SONICVOX_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "text": "Launch price $1,234.50, shipping 2026-08-30 at 14:30.",
    "voice_id": "cmpwrfi46015zkww3t8nui4ye",
    "apply_text_normalization": "on"
  }' \
  --output launch.wav

HTTP/1.1 200 OK
content-type: audio/wav
x-credits-charged: 100

<binary audio — "one thousand two hundred and thirty-four dollars
and fifty cents … August thirtieth, two thousand and twenty-six
at fourteen thirty">

Create a saved dictionary of respellings. Its alias entries apply automatically to every synthesis from your account — and, marked as the workspace glossary, to everyone in the workspace — merged UNDER any per-request lexicon ({ grapheme: respelling }, up to 200 entries, whole-word), which always wins; use_pronunciation_dictionaries: false skips the saved entries for one request. Alphabet honesty: only "alias" entries (plain respellings) are applied at synthesis. ipa, cmu-arpabet and x-sampa are accepted and stored but INERT — no deployed engine consumes phonetic alphabets, and this page says so rather than letting you ship a dictionary of dead phonemes.

Scope

voices:write

Credits

0

curl https://staging.sonicvox.ai/api/v1/pronunciation-dictionaries \
  -X POST \
  -H "sv-api-key: $SONICVOX_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "name": "Brand terms",
    "language": "en-US",
    "entries": [
      { "grapheme": "SonicVox", "phoneme": "sonic vox", "alphabet": "alias" }
    ]
  }'

HTTP/1.1 201 Created

{
  "dictionary": {
    "id": "cmf1x0000000000000000000",
    "name": "Brand terms",
    "description": null,
    "language": "en-US",
    "is_workspace_default": false,
    "entry_count": 1,
    "created_at": "2026-08-30T00:00:00.000Z",
    "updated_at": "2026-08-30T00:00:00.000Z"
  }
}