Cartesia TTS : Spanish voice Pronunciation bug, English accent applied to Spanish words

Support Help
1

For Cartesia TTS sonic 3.5 , we found that words are being mispronounced with an English accent, which sounds unnatural and confusing to native Latin American Spanish speakers. After investigating the issue directly through a Cartesia TTS API, we identified that a specific tag value is misconfigured, causing the pronunciation behavior to be incorrect.

For the language parameter, several values are available to represent different Spanish dialects. For Latin American Spanish, the recommended setting is “es-MX”.

“es” = generic Spanish (Cartesia picks a neutral accent)
“es-MX” = Mexican Spanish
“es-ES” = Castilian Spanish (Spain)
“es-US” = US Spanish / Spanglish accent

All of these language values produce the correct accent and pronunciation. However, when the language parameter is set to English (en), it results in the same incorrect pronunciation that we are currently observing in the Retell.

Could you please ask the team to look into this? Instead of setting the language parameter to en, can they update it to “es-MX” ? Based on our testing, using “es-MX” produces the correct Spanish pronunciation for Latin America, whereas “en” results in the same pronunciation issue.

Update this variable → language = “en” in the backend for cartesia config.

Screenshot from 2026-06-17 03-51-02
Screenshot from 2026-06-17 03-51-02311×209 9.45 KB

2

Hey @aamran Let me check with the team on this.

3

Hello @aamran the language picker is for STT, not TTS.

You need to choose a voice and voice model that matches the accent you want.

4

Screenshot from 2026-06-17 21-16-49
Screenshot from 2026-06-17 21-16-491451×422 54.4 KB

@Shaw Sorry about the confusion caused by previous image. When selecting the cartesia TTS voice in Retell, the voice traits indicate that it is Mexican, however, it does not consistently speak with a Mexican Spanish accent/pronunciation. Instead, it begins pronouncing words using an English accent and prounciation. If a word also exists in the English vocabulary, the system tends to prioritize the English pronunciation and accent.

We were able to reproduce this issue using the Cartesia API, when the language parameter was set to en.

Example Text:

Sí, quería quería contar de que tuve una buena jornada y, bueno, quería que me des algunos tips como para terminarla igual, de buena manera. ¿Eres un agente de asistente virtual

The pronunciation of “Jornada” and “virtual” is using English phonetics, with the J and TU sounds being pronounced according to English pronunciation rules.

The team can review the audio of above text from the Google Drive link, as I’m unable to upload audio samples here.

Alternatively, they can verify and reproduce the issue themselves by using the provided example and generating audio with the language parameter set to en or es-MX.

5

@aamran I have escalated this to the team.

6

Hey @aamran Can ou share the call ID?

7

Call Id = call_b8a1891ba1c87b190577b80ef0f

8

Hi @aamran Thanks for the Call ID, I have esclated this to the team for a review.

9

Any update on that ?

10

Hey @aamran For call_b8a1891ba1c87b190577b80ef0f, your agent is configured as single-locale, so we deterministically send language="es" to Cartesia on every TTS request (never "en"). The English-like pronunciation you’re hearing (“Jornada”, “virtual”) is how Cartesia’s cartesia-Alejandro + sonic-3.5 behaves under the neutral "es" hint.

Near-term options: try another LatAm/MX Cartesia voice; longer-term, we can file a request to emit region codes (e.g. es-MX) for es-419-only agents.

Thank You

11

@Shah-Fazal

we can file a request to emit region codes (e.g. es-MX) for es-419-only agents.

Can you assign that to this agent : agent_98123fc1c7d948c79729a9a165

With es-MX assigned to voice-id = custom_voice_36b2eddc930d8095593691254d

12

@aamran Escalated with the team

13

Please do notify me when it is done.

14

Hey @aamran Currently, we don’t have a per-agent override to emit es-MX
But Team have filed a feature request with engineering to add this.

15

@Shah-Fazal I’ve tested multiple voices from the cartesia library using different voice IDs, and they all exhibit the same accent issue.
I can consistently generate the correct Mexican Spanish accent on the Cartesia website, but when I use the same voice in Retell with a static sentence, the accent changes noticeably.

Is there a way for the team to configure or apply the es-MX locale/tag for those voice IDs in retell ?

Voice-Id :
5c5ad5e7-1020-476b-8b91-fdcbe9cc313c
15d0c2e2-8d29-44c3-be23-d585d5f154a1

16

@aamran Noted. I have forwarded your request to the team for review. We will update you as soon as we have more information.

17

Hey @aamran Team double checked our code and can confirm we are sending es language code to Cartesia in this case. Team also tested agent_98123fc1c7d948c79729a9a165 and could not hear any English accent in the voice, so we were unable to reproduce the issue. Would you be able to provide a sample recording? in Drive.

Thanks

18

Here is the link for one voice selected from Add Custom Voice → Community Voices:

Call-id for retell = call_ef250bf0ae1886eb6587c78260c

Here is the example text for the audio :

actualizar jornada virtual actualmente

For English speakers, Jornada is typically pronounced jor-NAH-duh, while Latin American Spanish speakers pronounce it hor-NAH-dah. Similarly, Virtual is pronounced VUR-choo-uhl in English, but beer-too-AHL in Latin American Spanish.

The issue we’re facing is that these certain word pronunciation in English accent, can confuse callers because the words sound different. As a result, callers often ask the AI agent to repeat the question, which disrupts the flow of the conversation and makes the interaction less smooth.

After testing, we found that certain % of words are pronounced using an English accent. While this percentage is small, it can disrupt the conversation.

Can you please ask the team if they can change the tag to es-MX for certain voices that appear under Add Custom Voice → Community Voices? This would allow us to easily identify and select the appropriate voices.

Voice below.

3597a26f-80ef-4bd5-8101-9699bc764917

15d0c2e2-8d29-44c3-be23-d585d5f154a1

b4b8e2af-6139-466e-a93a-30c20d2e1fc5

5c5ad5e7-1020-476b-8b91-fdcbe9cc313c

If the team updates the es-MX tag for only the four voice IDs mentioned above, we can test their behavior first. It’s possible that the changes could result in a higher percentage error of mispronunciations throughout the conversation, so we’d like to validate their performance.

19

@aamran Checking with the team.

20

@Shah-Fazal , any update on this ?