I’m using the assistant with ChatGPT 4.0, Deepgram (Hungarian), and 11Labs v2.5 (Hungarian voice) settings. The issue I’m encountering is that while the pronunciation is perfect at the start of the call, it becomes very unstable after a short period. The AI starts pronouncing Hungarian words with an English accent. The longer the conversation goes on, the worse the pronunciation becomes. I’ve noticed that the text generated by the AI during the call is not properly written in Hungarian, which seems to confuse the text-to-speech system and leads to pronunciation errors. This is problematic because users can immediately tell that they are not speaking to a human due to these mistakes. How can we resolve this issue to maintain the professional quality of the assistant?

@Broszki Moszki I'm sorry to hear that you're experiencing issues with the pronunciation stability during your calls. To address this, we should ensure that the language settings for both the speech-to-text (STT) and text-to-speech (TTS) systems are correctly configured for Hungarian. It's also important to check that the AI model is not defaulting to English during longer interactions. Here are the steps we can take to resolve the issue: 1. Verify that the language settings for Deepgram are set to Hungarian. If not, update the settings to use the correct language model. 2. Check the language configuration for the 11Labs TTS system to ensure it's set to Hungarian. 3. Review the AI model's language processing settings to confirm that it's not switching to English after a certain period or due to any triggers. 4. If the issue persists, we may need to look into the possibility of a bug or a memory issue that causes the model to default to English during extended conversations. If you have already checked these settings and the problem continues, it might be necessary to reach out to the support teams of Deepgram and 11Labs for further assistance. They may provide more insight into why the model's pronunciation is changing over time. Would you like me to guide you through checking and updating the language settings, or do you need assistance with contacting the support teams?

Please guide me through

> hey are not speaking to a human due to these mistakes. This is related to prompting, so work on prompting. > The AI starts pronouncing Hungarian words with an English accent. The longer the conversation goes on, the worse the pronunciation becomes. > For this can you share call id along with expected vs actual response with timestamp?

It appears to be working almost fine now after few tweeks in prompting. But,. Call id: 1ef7b894-dc9c-4322-9b83-bc2ab1828ca8 from 01:44 the pronunciation of the numbers are not good. Sounds like with English accent. Transcripts: User 0ad 328 3310. - (should be 06302083310) AI Teleszám, 0hat, 20, 3, egy 0od. Gépjárműű. Idái I30. - (should be "Telefonszám 06302083310. Gépjármű Hyundai i30") I think the main issue here is the speech to text whisper. I see lot of errors in the transcript.

hey @Broszki Moszki apology for the delayed response don't know how it got missed. Yeah your right you can try using deepgram and in your prompt you can instruct LLM to ask user to speak number digit by digit and then validate the same digit by digit this way it won't be sounding broken. Also you can ask LLM to add hypens

-

in b/w numbers to make it sound as you expected.