English heteronyms aren't pronounced properly

Checks

• This template is only for bug reports, usage problems go with 'Help Wanted'.
• I have thoroughly reviewed the project documentation but couldn't find information to solve my problem.
• I have searched for existing issues, including closed ones, and couldn't find a solution.
• I am using English to submit this issue to facilitate community communication.

Environment Details

I'm using voice cloning in audio.cpp docker with default settings.

I had Qwen3.8 go through the audio.cpp repo and check if it's anything in there. It believes the issue is Omnivoice. We also traced all the parameter flow and it looks like my project is using all the defaults.

Session make_request → generation_options_from_options(merged) fills OmniVoiceGenerationOptions with C++ struct defaults (types.h:19-36): seed nullopt, steps 32, guidance 2.0, speed 1.0, t_shift 0.1, denoise true, preprocess_prompt true, postprocess true, layer_penalty 5.0, position_temp 5.0, class_temp 0.0, chunk sizes, text_chunk_mode TagAware

Steps to Reproduce

"Read" and "live" are the big ones I keep noticing in normal usage. I just tested this in audio.cpp:

"What we must do is read. It's how we live. We watched the performance live, and yesterday we read books."

The pronunciations are hit or miss.