Hey everyone,I’m building a voice booking assistant for Client's Company using VAPI, and I’m running into a bunch of strange behaviour that doesn’t line up with my prompt or tool logic.
I’m hoping someone who has worked with complex VAPI assistants can help me diagnose what’s going on.
I have attached the VAPI prompt to this post, I have deleted my client's company from the prompt so if there are any gaps in tool names or text that is why
🔥 Issues I'm Experiencing:
1. The assistant sometimes lags badly
There are delays where the assistant just freezes for a second or two before responding.
2. It often says “I could not find that restaurant” even when the restaurant exists
The ID lookup tool works perfectly when I test it directly, but inside VAPI the assistant frequently fails to pick up names and jumps straight to “could not find that restaurant”.
It especially struggles with:
- “Alla Bolognese” → hears it as something completely different
- Restaurant names in general
3. The call audio breaks up / interrupts / talks over the user
It keeps speaking before I have a chance to confirm details.It will: - Interrupt me mid-sentence
- Repeat itself
- Apologise multiple times even though nothing actually happened
- Talk over the user during confirmations
It feels like endpointing or turn-taking is misfiring.
4. Repeats itself and seems “stuck” sometimes
At random points, it loops:
- Apologising repeatedly
- Repeating the same line
- Re-explaining a step it already completed
This doesn’t match the prompt at all.
5. Tone is inconsistent and sometimes shifts dramatically
This one is wild:
- It speaks in a normal tone
- Pauses
- Then suddenly shouts with excitement, like a different TTS voice style
- Then goes back to normal
It’s like ElevenLabs suddenly spikes in emotion even though stability is low and the prompt asks for calm, consistent tone.
6. Sometimes speaks in another language (!!)
While repeating an email address, the assistant suddenly switched to an incomprehensible language for a sentence before coming back.
Never seen this before.
7. Goes off-script or contradicts itself
Examples:
- I confirmed all details→ It then said it couldn’t find the place→ Then it cancelled the booking when I said “don’t cancel”
- It sometimes jumps forward in the flow before collecting required variables
- Sometimes it skips tool calls entirely and pretends it already ran them
8. Struggles with accurate speech recognition
Deepgram usually works fine for me, but here:
- It mishears restaurant names
- Mishears emails
- Mishears phone numbers
- Mishears dates/times
Even when spoken clearly.
🧩 What I’ve Tried
- Cleaned and simplified system prompt
- Reworked tool descriptions
- Checked all tool responses
- Adjusted cadence, endpointing, silence timeout
- Tested tools outside VAPI (they work perfectly)
- Tried with and without background noise
- Tried different voice models
Still seeing the same behaviour.
❓ What I'm Hoping to Understand
- Are others seeing stability issues with VAPI lately?
- Could this be a turn-taking / endpointing issue?
- Is GPT-4o inside VAPI known to “go off script”?
- Could the assistant be mixing messages due to overlapping tool calls?
- Are Deepgram transcripts sometimes causing chaotic behaviour?
- Is there something in my system prompt that could be conflicting with VAPI’s runtime?
💬 Any insight would be massively appreciated
This behaviour is extremely inconsistent and doesn’t match what the tool logic or prompt dictate — so I feel like I’m missing something deeper at the VAPI runtime level.
Happy to share any additional details if needed.
Thanks in advance!