Urgent: Persistent Hindi/Hinglish Language & Accent Issues with AI Assistant
Dear Vapi Support Team,
I am setting up a virtual receptionist, named Priya, for a premium dermatology clinic (Citrine Clinic). We are facing critical issues related to language switching and voice quality that are hindering the deployment.
Configuration Details:
Provider: OpenAI (using GPT-5/4o model)
TTS Provider: ElevenLabs
TTS Model Used: Eleven Multilingual v2 (Confirmed via config)
Voice Used: Monika Sogam (Confirmed via config to ensure native Indian accent)
Settings: Stability 0.5, Clarity 0.75
Prompt Type: Highly detailed, optimized hybrid prompt with explicit language enforcement rules (Hinglish examples).
The Problems:
Despite using the recommended multilingual model and a native Indian voice, two critical issues persist:
Language Detection Failure (The Loop): When the call connects, the assistant speaks the initial Hinglish greeting: "Namaste! Main Priya bol rahi hoon...". However, when the user immediately replies in Hindi/Hinglish, the AI often fails to maintain the language and defaults its next response to English.
Foreign Accent / Unnatural Voice: The voice, even with the Monika Sogam ID selected, reads basic Hindi words (like "Namaste" or "Jee") with an unnatural rhythm and a noticeable foreign accent. This suggests a possible conflict between the selected voice ID and the model rendering.
Steps Already Taken:
Verified the LLM (GPT) is receiving an explicit instruction to enforce language matching.
Confirmed the TTS Model is correctly set to Eleven Multilingual v2.
Confirmed the Voice ID is set to Monika Sogam (a known native Hindi voice).
Optimized prompt stability to reduce the robotic tone.
We suspect this is a platform-level issue related to how Vapi handles the language transition immediately following the initial fixed greeting, or a backend rendering issue with the chosen ElevenLabs voice and model pairing.
I am setting up a virtual receptionist, named Priya, for a premium dermatology clinic (Citrine Clinic). We are facing critical issues related to language switching and voice quality that are hindering the deployment.
Configuration Details:
Provider: OpenAI (using GPT-5/4o model)
TTS Provider: ElevenLabs
TTS Model Used: Eleven Multilingual v2 (Confirmed via config)
Voice Used: Monika Sogam (Confirmed via config to ensure native Indian accent)
Settings: Stability 0.5, Clarity 0.75
Prompt Type: Highly detailed, optimized hybrid prompt with explicit language enforcement rules (Hinglish examples).
The Problems:
Despite using the recommended multilingual model and a native Indian voice, two critical issues persist:
Language Detection Failure (The Loop): When the call connects, the assistant speaks the initial Hinglish greeting: "Namaste! Main Priya bol rahi hoon...". However, when the user immediately replies in Hindi/Hinglish, the AI often fails to maintain the language and defaults its next response to English.
Foreign Accent / Unnatural Voice: The voice, even with the Monika Sogam ID selected, reads basic Hindi words (like "Namaste" or "Jee") with an unnatural rhythm and a noticeable foreign accent. This suggests a possible conflict between the selected voice ID and the model rendering.
Steps Already Taken:
Verified the LLM (GPT) is receiving an explicit instruction to enforce language matching.
Confirmed the TTS Model is correctly set to Eleven Multilingual v2.
Confirmed the Voice ID is set to Monika Sogam (a known native Hindi voice).
Optimized prompt stability to reduce the robotic tone.
We suspect this is a platform-level issue related to how Vapi handles the language transition immediately following the initial fixed greeting, or a backend rendering issue with the chosen ElevenLabs voice and model pairing.
