Access by request

Scripts that read
like a person wrote them

Paste a script. It comes back with the machine-made phrasing gone, numbers and dates in the words you would actually say, technical and brand names marked so they are pronounced correctly, and formatted for whichever speech engine you are sending it to.

British English throughout, on a pronunciation dictionary of 2,058,954 words.

Request access Sign in

A work email address is required. Personal mailboxes are not accepted.

Why a model on its own gets this wrong

It changes your facts

Ask a model to tidy a script and it will quietly round a number, move a date or drop a name. Here every rewrite is checked against the original and thrown away if a fact went missing.

It cannot hear itself

A model has no idea which sounds a voice has barely produced. Coverage is measured against the 44 sounds of British English using a real pronunciation lexicon, not a guess.

It does not know the target

Every speech engine takes a different kind of hint. Send the wrong one and it reads the markup out loud as words. The output is shaped for the engine you actually picked.

One paste, six stages

What went in

In today's rapidly evolving landscape, we are delving into the transformative power of Kubernetes. We saw a 47% uplift on 3 March 2026, cutting costs by £12,500 and moving 2.3TB to Cloudflare R2, with Redis in front.

What came back

We saw a forty-seven percent uplift on the third of March, twenty twenty-six, cutting costs by twelve thousand five hundred pounds and moving two point three TB to /klˈaʊd flˈeə/ R2, with /ɹˈɛdɪs/ in front.

01

Slop

The tells that make writing sound machine-made are found and removed by rule, not by asking a model nicely. Deterministic, so the same script gives the same answer every time.

02

Rewrite

What is left goes through a rewrite that is checked afterwards. Every number, sum, percentage, date and proper noun has to survive. A rewrite that loses one is rejected and tried again, never shipped.

03

Numbers

Figures, dates, money and percentages become the words a person would actually say. The third of March, twenty twenty-six. Twelve thousand five hundred pounds.

04

Coverage

The script is measured against the 44 sounds of British English. Anything the voice will barely have said is reported, with extra lines to record if you are cloning a voice.

05

Pronunciation

Technical and brand terms are looked up and marked so they are said correctly. Redis is reddiss, not re-deez. Cidr is cyder. Your own names are saved and beat the dictionary.

06

Format

Pauses, tags and chunking are emitted in the form the engine you picked actually accepts, and split under its size limit.

“

A dictionary of headwords missed 18 per cent of a real script, because it has no inflected forms. A dictionary from 1913 then missed laptop, smartphone and sustainability, and guessed at them by splitting them in half. Both holes are filled.

CTC
Building the lexicon
2,058,954 British wordforms

Every engine, treated the same

Pick where the script is going and the output changes to match. Pause syntax, pronunciation syntax and the size it is split into all follow the engine, not a house preference. Limits below are each vendor's own documented figures.

EnginePause tagPronunciationLimit
Amazon Polly
6,000 characters per request, of which 3,000 are billed. SSML tags are not billed.
 Yes SSML, IPA 6,000
Cartesia Sonic 3
Takes a break tag but no pronunciation tag and no <speak> wrapper. No hard limit is published, so 2,000 is a safe working chunk; the Turbo variant caps at 500.
 Yes No 2,000
Deepgram Aura 2
Pronunciations go in as escaped JSON inside the text rather than as a tag. No pause tag, so timing comes from punctuation. Hard limit of 2,000 characters.
 No Inline JSON, IPA 2,000
ElevenLabs Flash v2.5
The only engine here that takes ARPAbet in a phoneme tag, and the largest limit of the set.
 Yes SSML, ARPAbet 40,000
ElevenLabs Multilingual v2
Takes pauses but no pronunciation tag, so unusual words are listed for you to check by ear.
 Yes No 10,000
ElevenLabs v3
No break tag. Pronunciation goes inline between slashes and pauses come from punctuation.
 No Inline IPA 5,000
Google Cloud Text-to-Speech
5,000 bytes per request, and the SSML markup counts towards it.
 Yes SSML, IPA 5,000
Hume Octave 2
No markup at all. Delivery is steered by a separate plain-English description field, so the script stays clean. 5,000 characters per utterance.
 No No 5,000
Inworld TTS
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
LMNT
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
Microsoft Azure AI Speech
Azure caps a request at ten minutes of audio rather than a character count. 8,000 is a safe working chunk.
 Yes SSML, IPA 8,000
MiniMax Audio
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
Murf
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
OpenAI gpt-4o-mini-tts
No markup of any kind. Delivery is steered by the separate instructions field, so the script itself stays clean.
 No No 4,096
OpenAI tts-1 and tts-1-hd
No markup and no instructions field. Everything has to be carried by the words themselves.
 No No 4,096
PlayAI
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
Resemble AI
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
Rime
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
Speechify
Markup support is not confirmed, so plain text is sent. 2,000 is a safe working chunk rather than a published limit.
 No No 2,000
Synthesia Express-Voice
Synthesia has run on its own Express-Voice engine since August 2025, with ElevenLabs voices available on Enterprise. Pronunciation is set in their editor rather than in the script, so plain text is sent.
 No No 2,000
Something else, plain text only
Safe default for any engine not listed. No markup is emitted at all.
 No No 5,000
Say it properly · Say it in British English · Say the numbers out loud · Say reddiss, not re-deez ·

Ask for access and a person will read it

Registration takes an email address and a one-time code. Some domains are let straight through. The rest are read by someone here, usually the same day.

Request access