Skip to content
Interface language: English
Audit / Signal

AI speech synthesis: the script is correct, but the narrator names another model

Check names, codes, measurements and sentence boundaries in a finished AI voice recording.

Article language: English

Four people stand around a circular pool; one touches the water.
In this article

The voice is calm, the sentences flow and music supports a film about workshop equipment. Everything seems right until someone notices that the narrator has repeatedly named a different product model. A zero has vanished from the code. Nobody heard a stumble: the error sounded perfectly confident.

This is a hypothetical production, but it shows what a speech-synthesis review must assess: the meaning of the finished audio. A correct script and an attractive voice are not enough if the listener hears another name, number or instruction. Test the recording itself and ask some questions without showing the text.

Decide what the listener must recognise

In our example, the fictional brand Nordvika presents an item coded N-08B with a dimension from 30 to 45 millimetres. The script contains three separate tasks. The listener must distinguish the brand from the model, reproduce the code without losing the zero or confusing a letter with a digit, and understand the two ends of the range with their unit.

The owner of the material should decide how the name is meant to sound. A model does not know the author's intention for a coined brand merely because the spelling is correct. Write the agreed pronunciation in words or keep a reference recording. If several pronunciations are acceptable, specify which still identify the same brand unambiguously.

For the code, decide how to read the letters using a convention agreed for the audience’s language. Do not leave this to chance, especially beside a foreign-language name. Mixing conventions can make a code hard to recognise even when the individual sounds are natural.

Do not treat zeros in an identifier as decorative. N-08B may identify something different from N-8B. Ask the listener to select the right version or write down the code they heard. Showing the script and asking "Did the narrator say it correctly?" is weaker: the listener may fill in a missing sound from the text.

For the range, establish the information to convey: from thirty to forty-five millimetres. The dash on the page is not itself an instruction for speech. A similar mark can occur in a code, a range or subtraction. Assess the meaning of the whole sentence, not just whether both numbers appear in the recording.

Distinguish a range from two dimensions. In this example, "30-45 mm" means a range, whereas "30 × 45 mm" can describe two sides. If the narrator reduces both to two numbers, the listener loses that relationship. The criterion must include what the numbers describe, not just their recognition.

Make a short list of elements whose pronunciation changes an identity or value. Here these are the name, code, range and unit. Elsewhere they might include an address, station number or company abbreviation. The list directs attention to information the listener needs; it need not cover every ordinary syllable.

A company glossary with a named owner can also hold agreed pronunciations. Someone must resolve uncertainty and approve changes. Otherwise each editor may replace the previous pronunciation with a personal preference, leaving the firm with several versions of the same product name. See the guide to creating a company glossary (in Polish) for a related perspective.

Separate script changes from pronunciation controls

A simple correction may expand an abbreviation in text intended for speech. In an appropriate sentence, "mm" can become the full name of the unit. This communicates intent but requires another listen. The spoken value must still match the number and unit in the approved material.

Do not unnecessarily change the text visible on the page. A phonetic spelling used to produce audio may look odd in a product caption. Separate approved display text from synthesis instructions while keeping their meaning aligned. The person reading the screen and the audio listener should recognise the same item.

Google Cloud describes selected SSML elements. SSML adds instructions to text for synthesis; Google documents support for a subset of the standard. Check the specific control in the product and voice you use instead of assuming every voice accepts the same settings.

Determine separately whether the tool can control code reading, name pronunciation and pauses between sentences. A speed slider can change the pace of the entire recording, but does not decide whether a code is read as a number, an abbreviation or a string of characters. Choose a correction that addresses the error found.

Microsoft describes custom pronunciation lexicons and language-dependent support. Such a dictionary can organise rules for repeated names. It is a preparation method, not evidence that the finished file sounds right; listen again.

Keep each rule with its context. A name may appear alone, inflected, at the start of a sentence or after another name. Check those uses if they occur in the material. A fix for a short label may not settle a different ending. The accepted result is a defined set of uses, not the mere existence of a dictionary.

Test a new rule against several previously correct passages. Expanding an abbreviation or changing how a string is read should not accidentally change an ordinary nearby word. Keep the example before the fix, the change and the expected result. This is more reliable than comparing two long recordings from memory.

Aeneia conceptual photography

If the tool lacks suitable controls, change the sentence intended for speech or use another solution for that passage. Agree on the route before producing many versions. Manual correction after export may be practical for one film and much harder across dozens.

Do not automatically transfer a rule from one language to another. The brand may stay the same, but surrounding words, inflection and code-reading conventions need checking. As with AI-assisted document translation (in Polish), consistent terminology is a separate criterion from fluency. The reviewer should know the language and intended meaning.

Listening without the script reveals what sight fills in

Give a sample to someone unfamiliar with the script. Ask them to identify the product, write the code and reproduce the dimension range. Do not warn them about the zero. Afterwards, show the text and compare differences. Record if several listens were needed instead of counting only the last successful attempt.

Choose a listener who resembles the intended audience. An employee who knows the catalogue may recognise the product despite a missing zero because they supply the gap from experience. A new customer may not. Record whether the listener had the catalogue, image or prior knowledge of the name. These conditions change what a correct answer proves.

Alternatives should test the actual distinction. N-08B versus N-8B tests the zero. A letter B versus a digit 8 tests another point. Do not offer only obviously different options, such as an item code beside the name of an unrelated category. An easy answer does not establish the distinction you need.

Listen to numbers as quantities. "Two and a half centimetres", "twenty-five millimetres" and "twenty-five centimetres" can have a similar rhythm, but the last quantity differs. If an equivalent unit conversion is acceptable, put that in the criterion. Do not automatically count it as an error or an improvement without comparing the quantities.

Check the complete sentence containing the condition, not only a clipped number. A pause or emphasis can separate a range from its unit or join the end of a code to the next fact. Listen through the transition, especially when separately generated clips have been joined. Fixing one word may change the rhythm next to it.

When a recording covers several products, ask the listener to pair each code with the correct name. Correctly reading each code alone is not enough if an editing pause joins one item's code to the next item's description. Two similar models and two different dimensions can test the required pairs.

The review of AI-generated video captions (in Polish) matters here too. Captions should match the finished voice and appear at the right time. If a pronunciation fix lengthens a passage, the old caption split may no longer fit. Do not approve the two independently when they must work together.

For names and codes in texts with different writing directions, also check character identity. Names, codes and addresses in bidirectional text (in Polish) concerns the visible layer. In synthesis, verify that the input order and spoken output identify the same item. The listener should reproduce the identifier without guessing its order.

Listen with the background that will remain in the published material. An ending that is clear in a clean file may disappear under music or a transition sound. Comparing the clean and final versions helps decide whether the synthesis or the edit needs correction. Do not ask a generator to fix a problem introduced during assembly.

Assess naturalness separately from accuracy. You can accept a name and still reject a clipped sentence ending, or like a voice and reject a wrong code. Separate these findings so the editor knows whether to change pronunciation, speed, editing or the text itself.

Accept the actual file and keep rules for the next one

After corrections, listen to the file intended for publication. A preview in the tool may differ from the edited film or the recording with music. Check beginnings, endings and joins. In the example, the first letter must not vanish with a trimmed breath, nor the final unit beneath fading music.

Keep the text, voice, language and relevant settings with the agreed pronunciation. Identify the final audio version and passages checked after changes. The next production can reuse the known rule without reconstructing it from memory, but a new name or generation method needs its own check.

The rules also matter when the voice changes. A new voice may have a better timbre but still needs the same list of names, codes and values checked. Keep a short comparison set with accepted answers. It need not cover the entire archive, but should include distinctions whose loss previously changed meaning.

If the recording is part of a conversation, a different question arises: when the assistant starts, stops and yields the floor. The voice assistant turn-taking review covers that scope. Good pronunciation of a finished line does not prove a sound interaction. Here the task is to accept the sound and meaning of an utterance the listener must recognise. For related guidance, see the guide to reviewing voice-assistant conversations (in Polish).

An AI-use review discussed with Aeneia Audit concerns the use of AI, data and rules. If you are considering a review of this use, agree its scope separately. Reviewing names and units is not voice identity testing or speech diagnosis.

In the final test, the listener should select N-08B and understand the dimension from 30 to 45 millimetres. If they identify another model, return to the discrepancy, however soothing the recording sounds. The product needs to remain the same when the listener closes their eyes.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear