A pleasant demo voice can still misread a customer's name, a delivery date or a product code. To choose text to speech for several languages, compare voices with the sentences your audience will actually hear. Keep a separate result for each language and accent.
This guide provides a test you can adapt for narration, support messages and product audio. The examples are test prompts, not measured results or a ranking of models.
Define the audience and the job
Write down the language, regional accent, channel and kind of message. Spanish narration for listeners in Mexico and Spain may need different voices. An English training video and a telephone greeting need different pacing.
Use the TTS language and accent directory to find a starting point. Check the voice available in Aisha Space. A language being available in the app does not establish which language codes or options its API accepts. Check those separately in the TTS API reference.
Build a small script that exposes mistakes
Use the same test cases for each candidate voice. Adapt the wording to each language with a fluent speaker. Literal translations can hide pronunciation problems.
| Test | Example input | What to listen for |
|---|---|---|
| Names | Please ask José to call Ms. Nguyen. | Both names remain recognizable. |
| Date and time | Your appointment is on 12 October at 14:30. | The spoken date and time match the intended meaning. |
| Product code | Your reference is AB-204. | Letters and digits are distinguishable. |
| Uzbek spelling | Bugun O'g'iloy bilan qo'ng'iroqlashamiz. | The reader preserves the names and Uzbek sounds. |
| Pauses | Your order is ready. Please bring your receipt. | The listener can hear the boundary between instructions. |
| Mixed terms | A sentence in your target language containing your brand name. | The voice handles the language switch without dropping words. |
Include ordinary sentences too. A set made entirely of difficult names will not show how the voice sounds through a full paragraph.
Make written text unambiguous
Decide what the listener should hear before generating audio. The date 03/04 can mean different things in different markets. Spell out the month. Decide whether 204 is a number or a sequence of digits in your reference code.
Keep the original text and the version prepared for speech. Record any expanded abbreviations or rewritten numbers. This shows whether an improvement came from the voice or from editing the input.
Listen with a scorecard
Ask a fluent listener to review each language. Hide the candidate names when comparing voices if practical. Use a simple scale such as 1 for unusable, 3 for usable after edits and 5 for ready for this specific task.
Score pronunciation, missing or added words, pacing and listening effort separately. Record the exact sentence and error beside every score. A wrong price or name should trigger a correction even if the voice sounds natural overall.
Regenerate difficult lines and check a full paragraph. One good sample does not establish consistent behavior. Keep the voice identifier, settings and evaluation date with the audio files.
Check the delivered audio
Listen through the channel your audience uses, such as a phone speaker or training platform. Check the beginning and end of the file for cut-off words. Confirm the output format and any downstream conversion your integration needs.
For an API integration, test one request, an invalid request and a retry using the current documentation. Measure response time on your own workload. Avoid extending the result from one language or sentence length to every supported voice.
Choose the next step
Keep the strongest voice for each language and list the sentences that still need editing. Retest those cases after changing a model, voice or text preparation rule.
Start a manual test in Aisha TTS, then use the API documentation when you need automation. If your starting point is a recording, follow the multilingual transcription guide before preparing text for speech.