34.【Voice AI Design】Safely Connecting an LLM

Why It Belongs Only in the Thinking State

tags: [“Voice Generation AI”, “LLM”, “FSM”, “Control Structure”, “Generative AI Design”]

🤖 Safely Connecting an LLM

— 🧩 Why It Belongs Only in the Thinking State

Through the previous articles, the voice AI has already been transformed into a stable system:

• 🧩 Fully managed by an FSM
• 🔊 Speaking is under strict control
• ✋ Interruptions no longer cause collapse

In other words, the skeleton is complete.

Only now is it safe to
bring the LLM back into the system.

🎯 Purpose of This Article

• Reconfirm where placing an LLM causes failure
• Determine where placing an LLM does not break the system
• Present a minimal and safe Voice × LLM coupling structure

👉 We are still not building a “chat AI.”
First, we decide on a placement that never causes accidents.

❌ Places Where an LLM Must NOT Be Used (Reconfirmed)

🔊 Speaking

• Strict real-time constraints
• Immediate stopping required
• Streaming control involved

👉 LLMs are too slow

👂 Listening

• Interruption decisions must be immediate
• Physical signals are sufficient

👉 Semantic understanding is unnecessary

✋ Interrupted

• Immediate transitions are critical
• Any delay leads to failure

👉 There is no room for an LLM

⭕ The Only Place Where an LLM Is Allowed

🧠 Thinking

The Thinking state:

• Has relaxed time constraints
• Produces no sound
• Is not interruption-sensitive
• Returns results as text

👉 A perfect match for LLM characteristics.

🧩 The Correct FSM × LLM Structure

🧩 FSM
 ├─ 👂 Listening
 ├─ 🧠 Thinking ──▶ 🤖 LLM (utterance content generation)
 ├─ 🔊 Speaking
 └─ ✋ Interrupted

Key principles:

• FSM → LLM is one-way
• The LLM never changes state
• The LLM never issues commands

👉 The LLM is a component, not a controller.

🔌 Information You May — and Must Not — Pass to the LLM

⭕ Safe to Pass

• The latest input text
• The purpose of the utterance (e.g., reply, confirmation, rejection)
• Maximum character length
• Tone constraints (e.g., neutral, formal)

❌ Never Pass

• State transition rules
• Interruption logic
• Timing control
• Continuation decisions

👉 The moment control information is passed, the system breaks.

🧪 Pseudocode (FSM × LLM)

if state == "Thinking":
    text = llm_generate(prompt)
    state = "Speaking"
    play_tts(text)

Key points:

• The LLM is called only once per Thinking state
• Its output is just a string
• The FSM alone decides the next state

🚫 Common and Dangerous Anti-Patterns

❌ Asking the LLM “What should we do next?”

→ State management collapses

❌ Asking the LLM “Should I keep speaking?”

→ Infinite speech

❌ Letting the LLM decide interruptions

→ Immediate failure

🧠 Why This Structure Is Stable

• The FSM controls time
• The FSM controls input and output
• The FSM absorbs exceptions
• The LLM generates meaning only

👉 Responsibility separation is complete.

📌 Summary

• 🤖 Place the LLM only in the Thinking state
• 🧩 The FSM always retains control
• 🔊 Never put an LLM in Speaking
• ✋ Never use an LLM for interruption decisions
• 🏗 The LLM is “external intelligence,” not the core

🏁 Series Conclusion

In this series, we fixed the structure of voice AI by establishing that:

• 🎧 Voice generation AI is a control problem, not a generation problem
• 🧩 Systems centered on FSMs do not collapse
• ✋ Interruptions are normal behavior, not exceptions
• 🤖 LLMs must be confined strictly to the Thinking state

Even so, many people will still say:

🎭 “I want it to feel more conversational”
🤝 “I want it to feel like talking to a human”

But that is not a design objective — it is a desire for performance.

The minimum requirement for a voice AI to appear intelligent
is not imitation of conversation.

It is simply this:

• States are explicit
• Transitions are predictable
• The system can always recover from failure

That is control integrity.

This series ends by
breaking the illusion of conversational AI.

Anything beyond this
can be explored later — if and when it is truly needed.

(The GitHub-hosted Markdown is the canonical source; the Qiita version is a curated extract.)