I’m experiencing a consistent issue with Peninsular Spanish voices from ElevenLabs in Retell AI (Turbo 2.5 model) that’s affecting my users’ experience.
Problem details:
I’m using Peninsular Spanish voices exclusively (single-language configuration)
The voice “freezes” unpredictably during pronunciation
When it happens, words are pronounced incorrectly or oddly, as if speaking in a different language
Duration varies: from issues with specific words to 2+ second inconsistencies
I’d like to understand:
Why is this happening?
Is it possible to fix it?
What’s the expected timeline for a solution?
Are there any actions I can take on my end to mitigate or prevent it?
I appreciate any guidance or updates on this matter.
Sorry for the late answer. I’m going to send another call ID where clearly the Spanish voice is freezing or being distorsioned. It’s really affecting the experience.
Are there any potential solutions to this issue?
I’m currently using the turbo model and the call ID is: call_a05842b3363b6d3ad72f833aab0
Hey @salah.eddin.an Team pulled call_a05842b3363b6d3ad72f833aab0 and audited the full pipeline. The “freeze” is not a delivery issue on our side. Audio streamed normally the whole call with no synthesis errors, no disconnects, and no fallback to your Cartesia voice.
The cause is in the content sent to TTS. Your Step 1 greeting writes the brand name phonetically in Spanish (treiding traks and machíneri). On this call, machíneri took 639 ms to render, roughly 2x any other word, with a pause after it. That is the “freeze.” Turbo V2.5 is rendering an English brand name spelled in Spanish phonemes, which stretches the token and inserts an unexpected pause. Your voicemail text on the same agent uses the English spelling “Trading Trucks and Machinery” and renders fine, which confirms it is the phonetic rewrite, not the voice.
What to change, in priority order:
Rewrite the greeting with standard English spelling: Trading Trucks and Machinery. Biggest single fix.
Switch the voice model from eleven_turbo_v2_5 to eleven_multilingual_v2. It holds cloned-voice identity better in Spanish and is the tier where Pronunciation Dictionary and <phoneme> SSML actually work (V2.5 ignores them). Expect a small latency increase for steadier prosody.
Lower voiceTemperature from 0.88 to ~0.5 to 0.6.
There is no platform-side fix on a timeline because the runtime is working as designed here. All three changes are on the agent config side and you can apply and test them today.