Feature Proposal: “Hold Music” — Expectation-Setting Voice Prompts During High Latency

[rosscado]

During long response times (e.g. 10+ seconds when GPT is “thinking”), users currently sit in silence until the first audio stream begins. This can feel awkward or like the system hasn’t heard them.

Instead of filler sounds (“um,” “ah”), we can use this gap to set expectations in a natural, human way. Example: a short cached line in the selected voice such as:

• “Got it, let me think about that for a moment.”
• “Hmm, that’s a big one. Give me a second.”
• “Okay, I’m working on it.”

This reassures the user they’ve been heard and prepares them for a thoughtful response, similar to how in human conversation it’s acceptable to say, “Let me think about that.”

---

Proposed Approach

1. Latency detection

   • Detect when a model’s expected response time > ~10s (e.g. GPT-5 thinking mode).
   • Only trigger “Hold Music” prompts in these cases, not for instant responses.

2. Prompt playback

   • Short audio snippets in the selected voice + language.
   • Can be cached phrases or generated on the fly (TTS).
   • Randomize from a small pool for variety.

3. User experience goals

   • Communicate “you’ve been heard.”
   • Reduce awkward silence.
   • Maintain tone consistency (voice, language, pacing).
   • Feel natural and optional (can be toggled off).

---

Technical Options

• Cached audio prompts per supported voice → cheap, fast, reliable.
• Dynamic generation with lightweight LLM → flexible, more variety, but more expensive.

Fallback: If no audio asset is available in the current voice/language, skip playback.

---

Open Questions

• Threshold: is 10s the right cutoff? Should it be adaptive?
• Should this live entirely client-side (pre-recorded prompts) or server-side (generated per request)?
• Do we expose a user setting: “Play expectation-setting prompts when response time is high”?
• How do we avoid confusion between “Hold Music” and actual AI content (make it clearly a system/acknowledgement message)?

---

Why It Matters

• Makes conversations feel smoother and more human.
• Sets the right expectations instead of leaving users in silence.
• Reinforces trust that the assistant is listening, even when it needs time to think.

[rosscado]

Context: Lessons from #Keep4o and #SaveStandardVoice

Recent user movements around model and voice changes highlight that people value presence, patience, and trust as much as raw speed or technical benchmarks.

• #Keep4o (Aug 2025): Users protested when GPT-4o disappeared. The reaction wasn’t about benchmarks — it was about losing the warmth and reassurance they’d grown attached to.
• #SaveStandardVoice (Sep 2025): Users resisted the retirement of Standard Voice Mode because it offered full, thoughtful answers in calm, steady voices, whereas Advanced Voice felt “rushed” or “perky assistant-like”.

Both cases show: silence, abruptness, or clipped responses are experienced not just as UX flaws, but as breaks in relationship. Users want AI that signals it is listening and taking their input seriously, even if that means waiting longer.

The proposed “Hold Music” expectation-setting prompts align with this lesson:

• They preserve continuity of presence during latency.
• They echo how humans naturally manage silence in conversation (“let me think about that”).
• They give users trust and comfort that the system is attentive, not frozen.

In other words, this is about more than smoothing latency. It’s about designing for relational continuity — keeping the sense of being heard, even when the system needs time to respond.