Voice Live (gpt-realtime) content_filter cuts harmless audio mid-sentence – much higher rate with children's voices

We are building an English speaking-practice app for elementary-school children (Taiwan), using the Voice Live API with the fully managed gpt-realtime model (region: East US 2, api-version 2026-04-10, voice: shimmer). We have no gpt-realtime deployment of our own. All model output is short, ordinary teaching language.

Problem

Some responses end with status incomplete and status_details.reason: "content_filter". In every case the text transcript is complete and harmless, but the audio stops partway through. Examples of responses that were cut:

• "Repeat after me: Can I get some coleslaw?"
• "Great work, that's clear! Number 4! Look at the card and tell me in English."
• "Let's try in English, 好嗎?"
• "Nice work, very clear! Great job today! Let's look at what we learned."
• "Great try, just a little off. Let's fix the order of words. Repeat after me: How much is it?"

Observations

• With adult voices or text input, about 1% of responses were cut (8 of 742).
• With a child speaking, it was 12% (6 of 49, two sessions).
• In the child's sessions, when the same turn was requested again in the same conversation, it was cut again. In adult sessions the retry succeeded.
• Cut responses contain much more audio than normal for the same sentence (about 155 vs 63 audio tokens), and the extra part is not speech.

I can provide response IDs and UTC timestamps privately if that helps.

Questions

1. Is there a way to find out which content filter category and severity triggered these interruptions? The event only reports content_filter.
2. Is the filter evaluated on the generated audio separately from the text? Since the text is harmless, we suspect the abnormal audio is what gets flagged.
3. Is there a known issue where a child's voice in the input audio makes the output more likely to be filtered? Is there a recommended mitigation (session settings, response length, etc.)?
4. The Voice Live FAQ says the content filtering cannot be modified. Is bring-your-own-model the only way to apply a custom content filter, and does it fully replace the built-in Voice Live filter? Our earlier quota request for a gpt-realtime deployment was denied; what is the right way to request it for an educational use case?

Thank you,

Hao Chun Liao