Text normalization

Rime automatically normalizes English text — numbers, dates, currency, phone numbers, measurements, and more — before synthesis.

When you send text to Rime’s TTS models, a normalization layer runs first. It expands numbers, dates, currency, phone numbers, measurements, and other non-standard words into their spoken form before the model synthesizes audio.

Rime handles text normalization automatically. Most common formats — currency with symbols, dates with years, clock times, phone numbers, and standard measurements — expand correctly without any preprocessing. Just write naturally. If something sounds wrong, debug it with /textnorm before adding a pre-processing layer to your application.

What’s handled, at a glance

Click any category for the full input/output reference.

Category	Examples	Reference
Numbers, currency, ranges, measurements	$1,045.96, 5kg, 98°F, 13-50, 1/2, 1e6, [phone]	Numbers, currency, and measurements
Dates and times	10/12/2024, 2021-03-15, April 2, 2024, 3:45pm, 15:45, noon	Dates and times
Addresses, URLs, emails	529 Main St., Boston, MA 02129, https://app.rime.ai, [e-mail]	Addresses, URLs, and emails
Abbreviations, acronyms, initialisms	Dr. Smith, e.g., NASA, d. n. a.	Abbreviations, acronyms, and initialisms
Symbols and percentages	&, $, %, 100%	Symbols and percentages
Punctuation and prosody	,, ., ?, ...	Punctuation

Forced letter-by-letter reading

For account numbers, confirmation codes, SKUs, and acronyms the normalizer doesn’t recognize, wrap the string in spell(...) to force letter-by-letter pronunciation. Works on both Arcana and Mist.

Input:  Your confirmation code is spell(PRM423GDDML2354).
Output: Your confirmation code is P R M, 4 2 3, G D D, M L, 2 3, 5 4.

For full reference, see Spell function.

Brand names, product names, and uncommon words

Rime’s models may not nail uncommon brand or product names on the first try. Two options:

1. Submit the word to Rime to add to the dictionary (around 24 hours). The fastest way is the Speech QA dashboard, which surfaces out-of-vocabulary words from your traffic and lets you approve or correct them in one place. You can also reach the team directly at [e-mail] or via your shared Slack channel.
2. Use custom pronunciations inline with Rime’s phonetic alphabet and phonemizeBetweenBrackets: true. See Custom pronunciation for the full reference.

phonemizeBetweenBrackets works on the Mist family (v1, v2, v3). It does not work on Arcana. For brand or product name pronunciations on Arcana, submit the word to Rime to add to the dictionary, respell phonetically in plain English (accepting that this is approximate), or use Mist for flows where pronunciation control matters.

For full reference, see Custom pronunciation. To check whether a word is already in Rime’s dictionary, use the Coverage API.

Feature availability: Arcana vs. Mist

The features below apply to the Mist family (v1, v2, v3) unless otherwise noted.

Feature	Arcana	Mist
Native text normalization (numbers, currency, dates, etc.)	✅	✅
spell() for forced letter-by-letter	✅	✅
Punctuation-driven prosody	✅	✅
pauseBetweenBrackets (custom pause tags like <750>)	❌	✅
phonemizeBetweenBrackets (inline phonetic strings)	❌	✅
Deterministic per-term pronunciation config	❌	✅

Arcana has parity with Mist for numbers, currency, and abbreviation expansion. If you tested Arcana early and saw normalization issues, re-test on the current model.

For flows that need precise pause durations (legal disclaimers, regulated read-backs) or guaranteed pronunciation of brand and product names, Mist is the safer choice.

Debugging with the textnorm endpoint

Rime exposes a /textnorm endpoint that returns the normalized form of an input string — exactly what the TTS model receives before synthesis. It’s the fastest way to separate normalization issues from synthesis issues.

curl -X POST https://optimize.rime.ai/textnorm \
    -H "Authorization: Bearer $(rime key)" \
    -H "Content-Type: application/json" \
    -d '{"text":"1234 1,2,3,4 1-800-444-4141 "}'

{"normalized":"one two three four, one , two , three , four, one, eight hundred, four four four, four one four one"}

This endpoint covers Rime’s English text normalization. Output is the same regardless of which model you’ll synthesize with. For the full request and response reference, see the Text Normalization API.

Triage workflow

When something sounds off:

1. Capture the exact input text that produced the bad output.
2. POST it to /textnorm and look at the normalized output.
3. Compare the normalized output to what you expected the model to say.
4. If normalization is wrong, you have a reproducible signal — flag it to Rime with the input, expected normalization, and actual normalization. Fixes ship on Rime’s side.
5. If normalization looks correct but speech still sounds off, the issue is in synthesis, not normalization — try a different voice, model version, or sampling settings.

Testing checklist

Before going to production, test the voice against realistic versions of:

1. Every date format your backend can produce (MM/DD, MM/DD/YYYY, ISO, relative like “tomorrow”).
2. Every currency you’ll quote, including round numbers and fractional cents.
3. The longest realistic phone number, account number, and confirmation code.
4. Your top 20 most-spoken product or brand names.
5. At least one utterance from each of: quote, greeting, confirmation, error, payment, scheduling.
6. The same utterance regenerated 5 to 10 times. Consistency across regenerations is what catches sampling variance.

For anything that sounds wrong, POST the exact input to /textnorm to see what the model actually received, then flag it to Rime so the fix lands for everyone.

Pre-normalizing in your application

For most applications, pre-normalizing is unnecessary and adds latency and engineering complexity. Rime’s normalizer handles common patterns natively, and the fastest way to fix a pronunciation issue is to verify with /textnorm and flag any miss to Rime.

If you have a specific pattern that genuinely needs to be pre-expanded — for example, an alphanumeric ID format unique to your domain, or a flow where regenerated text must read identically every time — see Pre-normalizing text for guidance and a drop-in prompt template.