Preparing Text Before a Speech Engine Reads It

Speech engines read ordinary sentences well. They fail on the parts that are not
sentences: figures, abbreviations, links, addresses. The failures are quiet — nothing
errors, you just get a file with a wrong word in the middle of it — and you only find
them by listening to the whole thing.

Ten minutes of preparation removes almost all of them.

Numbers are the main event

A bare figure forces the engine to guess how to say it, and the guess depends on
surrounding context in ways you cannot predict. In our own testing a four-digit number
followed by a hyphenated word came out as something that was not a number at all.

Writing it out in words removes the guess. 5070 becomes “five thousand
seventy” and there is nothing left to interpret. This is the single highest-value edit you
can make to a script.

Thousands and decimals are language-specific, and they are opposites

This trips people up constantly, because the two conventions are mirror images:

• English: 1,180 is one thousand one hundred eighty;
  3.14 is a decimal.
• Turkish and much of Europe: 1.180 is one thousand one
  hundred eighty; 3,14 is a decimal.

Read a European price with English rules and a figure meaning three and a bit becomes three
hundred and fourteen. The two languages also say decimals differently: English reads
the digits one at a time (“three point one four”), while Turkish reads the fraction as a whole
number.

Leave codes alone

Not every digit should be converted. MP3, H264 and model numbers
are read correctly as they are, and spelling them out makes things worse. The rule that works:
convert digits that stand alone, leave digits attached to letters.

The rest of the checklist

• URLs and e-mail addresses. Read aloud, they are gibberish, and every
  character is billed. Replace them with a short phrase or remove them.
• Repeated whitespace and stray line breaks. Harmless to meaning, but they
  count toward your quota and can affect phrasing.
• Curly quotes and long dashes. Normalise them. Some engines pause oddly
  on typographic punctuation.
• Bracketed notes and stage directions. If it is not meant to be spoken,
  take it out — it will be spoken.

Split before you send, not after it fails

Most engines cap a single request at around 5,000 characters. A long chapter has to be
split, and where you split matters: break on sentence boundaries, never mid-sentence, or the
join will be audible.

Knowing the request count in advance is also how you plan the cost. Speech is billed per
character, so the character count of your cleaned text is the price. Clean first,
count second, and you know what an audiobook chapter costs before you commit to it.

How long will it actually be?

Words divided by a speaking rate. 180 words per minute is a typical narration pace;
audiobooks usually run slower, conversational podcasts faster. A 9,000-word chapter is
roughly 50 minutes at a normal pace — useful to know before you find out the hard way
that it is too long for one episode.

Readability, and one caveat

Readability scores are worth a glance, with a warning: the well-known formulas are
calibrated per language. Flesch Reading Ease is built on English syllable patterns and returns
a meaningless figure for an agglutinative language like Turkish, which has its own adaptation
(the Ateşman index). A tool that applies the English formula to any text is giving you a number
that looks authoritative and means nothing.

Why do this in the browser

Scripts are often unpublished manuscripts, client work or internal documents. Text
preparation is pure string manipulation — there is no reason for any of it to leave your
machine. If a tool uploads your draft to a server to count its characters, that is a choice it
made for its own convenience, not a technical requirement.

The order that works

1. Paste the text and look at the character count.
2. Convert numbers to words in the right language.
3. Clean URLs, addresses and stray whitespace.
4. Check the request count and the longest segment.
5. Send it to synthesis.
6. Master the result before publishing.

Steps two and three take a minute and remove the class of error that is hardest to find
later.