🎓 How to Train AI Voice Bot in Nepali, Hindi, or English (2026 Guide)

You Don't "Train" — You Configure

Modern AI voice bots (powered by Gemini Live, GPT-4o, etc.) don't require ML training. The model is already trained on massive datasets. Your job is to configure it for your specific use case via:

1. System prompt — personality, rules, conversation flow
2. Knowledge base — facts about your business
3. Examples — sample conversations showing desired behavior
4. Voice/language settings — which voice, which language

Structuring Your System Prompt

A well-structured Nepali voice bot prompt has these sections:

1. Identity (Who Are You?)

तिमी TalkC AI हौ — XYZ Company को आधिकारिक भ्वाइस एजेन्ट।
तिमी मानिस जस्तै कुरा गर — robot होइन, मानिस।

2. Language Rules

- Default: नेपाली
- Caller ले English बोले: English मा switch गर
- Caller ले Hindi बोले: Hindi मा reply (तर तिम्रो default नेपाली नै हो)
- नेपाली-English mix ("balance check garnu") = normal नेपाली, language switch नगर

3. Tone and Style

- हजुर, हस्स, हुन्छ नि — natural Nepali fillers use गर
- Caller को tone match गर — angry → empathetic, happy → cheerful
- SHORT reply — 1-2 sentence then STOP, let caller talk
- Repetition नगर — same phrase दुई पटक नभन

4. Conversation Flow

═══ GREETING ═══
"नमस्ते हजुर! XYZ Company मा स्वागत छ। कसरी मद्दत गरौं?"

═══ ANSWER STYLE ═══
- Caller ले प्रश्न सोध्यो? सिधा उत्तर देऊ
- Answer दिसकेपछि stop — "अरु केही?" बारम्बार नसोध

═══ ESCALATION ═══
- "मान्छेसँग कुरा गर्न चाहन्छु" सुनियो? → human transfer
- 2 पटक try गरेर solve भएन? → escalate

5. Knowledge Base Section

═══ TIMRO COMPANY ═══
- Name: XYZ Restaurant
- Hours: 11 AM - 10 PM daily
- Address: Thamel, Kathmandu
- Phone: 01-1234567
- Speciality: Newari thali

═══ MENU ═══
- Veg Thali: Rs 350
- Newari Khaja Set: Rs 550
- Mo:mo (10 pcs): Rs 150
- ...

Hindi-Specific Considerations

Honorifics

• आप (aap) — formal, default for AI
• तुम (tum) — informal, with younger callers
• तू (tu) — very informal, avoid

Common Hindi Code-Switching

• "App download कैसे करें?" — natural mix, respond similarly
• "Balance check karna hai" — respond in mixed Hindi
• Caller English wholly → switch to English

English Configuration

English is "easier" for AI but has its own challenges in South Asia:

• Indian/Nepali English accents — choose STT optimized for these
• Regional vocabulary — "Doing the needful," "Kindly revert," "Prepone the meeting"
• Formality level — South Asian English tends more formal than American

Knowledge Base Best Practices

Structure Information Hierarchically

• Company overview (1 paragraph)
• Core services/products (bullet list)
• FAQs (Q&A format)
• Pricing rules (when AI can share vs escalate)
• Common scenarios (with sample responses)

Keep It Conversational

Write in the language and style the AI should respond. Don't paste technical documents:

• ❌ "Our enterprise SaaS solution leverages cutting-edge..."
• ✅ "We help businesses save time by automating phone calls."

Include Real Phrases

If callers ask things in specific ways, include those phrases:

• "How much does it cost?"
• "Kati paisa lagcha?"
• "Price kya hai?"

And paste the answer in same language nearby.

Testing Strategy

1. Internal Testing (Day 1-2)

• Make 20+ test calls
• Test edge cases: long pauses, interruptions, accents
• Test multilingual: switch mid-call
• Test escalation: ask for human transfer

2. Soft Launch (Day 3-7)

• Route only off-hours calls to AI initially
• Review every recording
• Identify patterns of misunderstanding
• Refine prompt

3. Full Launch (Day 8+)

• All calls routed to AI
• Monitor dashboard daily
• Weekly review of sentiment + tickets
• Iterate on knowledge base monthly

Common Mistakes to Avoid

1. Prompts That Are Too Long

5000+ token prompts cause latency. Keep under 3000 tokens. Use knowledge base for facts, prompt for behavior.

2. Translation Instead of Native

Don't write in English then translate. Write directly in the target language. Awkward phrasing kills user experience.

3. No Anti-Repetition Rules

Without explicit rules, AI will repeat phrases like "अरु केही?" every turn. Always add: "Same phrase दुई पटक नभन।"

4. No Voice/Personality

Generic "AI assistant" prompts produce generic AI. Add personality: warm/professional/playful — match your brand.

5. Missing Escalation Rules

AI will try to handle everything if you don't tell it when to give up. Add specific triggers for human transfer.

Advanced: Function/Tool Calling

For dynamic data (real-time order status, appointment booking), define tools the AI can call:

{
  "name": "check_order_status",
  "description": "Look up order by order number",
  "parameters": {
    "order_number": "string"
  }
}

AI decides when to call the function based on conversation context.

Continuous Improvement

• Weekly: Review 10 random call recordings, note issues
• Monthly: Update knowledge base with new FAQs, pricing changes
• Quarterly: Major prompt refactor based on patterns
• Annually: Consider switching to newer model if quality improves

Frequently Asked Questions

Can I train AI voice bot myself or do I need an expert?

For basic deployments, anyone with good writing skills can configure an AI voice bot. For complex use cases (tool calling, multi-step workflows), you may benefit from prompt engineering experience.

How long is a good system prompt?

1500-3000 tokens (roughly 1000-2000 words). Longer prompts slow response time. Keep behavior rules in prompt, push facts to knowledge base.

Should the prompt be in Nepali/Hindi or English?

Write prompts in the target language for best results. Mixed prompts (English instructions, Nepali examples) also work. Avoid translating English prompts — write natively.

How do I handle multiple languages simultaneously?

Modern voice models handle code-switching naturally. Set primary language, allow switching based on caller. Test mid-call language switches explicitly.

Can AI learn from each call automatically?

Not in real-time. You manually update prompts/KB based on insights. Some platforms offer call analytics to identify improvement opportunities. True online learning is risky for production voice bots.