We’re running a voicebot product built on top of the OpenAI Realtime API (currently using gpt-realtime-mini). We have noticed that the Ballad voice is changing tone mid-conversation or between calls — shifting from its expected calm/warm style to something more neutral or flat, without any change on our end.
What we’re seeing:
-
Same system prompt, same session config
-
Ballad sounds noticeably different across calls — sometimes warm/expressive, sometimes flat
-
Other voices (e.g. Alloy, Echo) seem more consistent in the same setup
Setup:
-
Realtime API via SIP integration
-
Sessions initiated programmatically with identical parameters
-
No changes to voice, instructions, or turn_detection config between sessions
Questions:
-
Has anyone else noticed tone/personality drift in the Ballad voice specifically?
-
Is there any known variability in voice characteristics tied to model updates or infrastructure changes on OpenAI’s side?
-
Is there a way to “anchor” voice personality more reliably through the system prompt?
We’re considering removing Ballad from our available voice options for this use case, but wanted to check if this is a known issue or something with a workaround before doing so.
Any input appreciated.
Absolutely, seconded 
. I came here to report this problem.
Ballad’s voice on v1 was very enthustiastic - possibly a little too enthusiastic - but very reliable. On v1.5 it became much more calm and an upgrade on v1, but reliability became an issue. Sometimes ballad’s voice breaks, he will suddenly pitch up or even sound female. Sometimes he can be all over the place. v2 the problem persists.
We are a British firm and all the other voices are American accents so Ballad is our preference but we have switched to Marin as default for now.
Thirded here 
Also UK based so Ballad was our go-to option for voices, but having real issues with using any of the available models. Pitch constantly changing is extreme ways.
Thanks for the Marin suggestion, this is working well with an accent part of the prompt like this:
## Accent
Speak English with a British accent.
- Keep the accent stable from the first word to the last.
- Use natural British vowel shaping, but keep speech easy to understand.
- Do not exaggerate the accent.
- Do not change response language based on the user's accent.
What’s interesting is that this works reliably when using realtime-mini but has issues with accent consistency when using realtime, realtime1.5 or realtime2. With it not working at all, or reverting to a US accent after tool usage.
The lack of feedback from OpenAI on this is pretty frustrating, there a set of threads with this being discussed over the last couple of months. This was not originally present when we built PoC’s back at the start of the year.
Thanks for the detail, and sorry for the delay. We understand the voice and accent consistency issues, including after tool calls.
We are going to close this topic. If the issue continues, please create a new topic with current details and share the exact Realtime model, session settings and a short sample privately with Support: https://help.openai.com/en/articles/6614161.so it can be investigated.