Build a Malayalam voice agent on Bolna

​

Choose a Malayalam transcriber

​

Choose a Malayalam voice

​

Write the prompt for Malayalam

• Write in Malayalam script, not romanized Malayalam. നമസ്കാരം! സുഖമാണോ? rather than Namaskaram! Sukhamano?. Native script gives more accurate pronunciation and stops the text being treated as English.
• Keep common English loanwords in Latin script. Callers say “order”, “OTP” and “support ticket” in English, so keep those words as they are inside Malayalam script text instead of forcing a pure Malayalam translation.
• Set gender and politeness explicitly. These are often carried by verb forms, so name the forms you want for the agent’s persona rather than writing “be polite”.
• Read alphanumerics character by character, in English. For OTPs, registration numbers and order IDs, this avoids misheard and mispronounced digits.
• Leave language switching out of the prompt. Bolna switches languages itself.

​

Create a Malayalam agent over the API

{
  "agent_config": {
    "agent_name": "Malayalam Support Agent",
    "agent_welcome_message": "നമസ്കാരം! എനിക്ക് നിങ്ങളെ എങ്ങനെ സഹായിക്കാനാകും?",
    "tasks": [{
      "task_type": "conversation",
      "toolchain": { "execution": "sequential", "pipelines": [["transcriber", "llm", "synthesizer"]] },
      "tools_config": {
        "llm_agent": {
          "agent_type": "simple_llm_agent",
          "agent_flow_type": "streaming",
          "llm_config": { "provider": "openai", "model": "gpt-4.1-mini", "max_tokens": 150, "temperature": 0.2 }
        },
        "transcriber": {
          "provider": "sarvam",
          "model": "saaras:v4",
          "language": "ml",
          "stream": true,
          "endpointing": 500,
          "keywords": "Bolna, PhonePe, Flipkart",
          "context": "Inbound support line for an electronics retailer in India. Callers ask about order status and returns, and often read out order numbers."
        },
        "synthesizer": {
          "provider": "sarvam",
          "provider_config": {
            "voice": "Priya",
            "voice_id": "priya",
            "model": "bulbul:v3",
            "language": "ml",
            "speed": 1.0
          },
          "stream": true,
          "buffer_size": 150
        },
        "input": { "provider": "plivo", "format": "wav" },
        "output": { "provider": "plivo", "format": "wav" }
      },
      "task_config": { "call_terminate": 90, "hangup_after_silence": 10 }
    }]
  },
  "agent_prompts": {
    "task_1": {
      "system_prompt": "നിങ്ങൾ ഒരു സഹായകരമായ സപ്പോർട്ട് ഏജന്റ് ആണ്. ചുരുക്കി മറുപടി നൽകുക."
    }
  }
}

curl https://api.bolna.ai/v2/agent \
  -H "Authorization: Bearer $BOLNA_API_KEY" \
  -H "Content-Type: application/json" \
  -d @agent.json

import os, urllib.request, json

with open("agent.json") as f:
    body = json.load(f)

req = urllib.request.Request(
    "https://api.bolna.ai/v2/agent",
    data=json.dumps(body).encode(),
    headers={"Authorization": f"Bearer {os.environ['BOLNA_API_KEY']}",
             "Content-Type": "application/json"},
)
print(json.load(urllib.request.urlopen(req)))

​

Improve recognition with keywords and context

• keywords: a comma-separated list of terms callers say.
• context: a sentence or two describing the call.

• Sarvam: keywords on Sarvam needs saaras:v4, and takes up to 50 terms of up to 64 characters each. context works on every Saaras model.
• Soniox: reads both. context can run to about 10,000 characters.
• Smallest AI: reads keywords, with an optional weight per term such as Bolna:5, up to 100 terms. It has no context field.
• Gemini is not in the provider support table for these fields.

​

Fix pronunciation of numbers and names

​

Serve Malayalam and English on one agent

• Code-mixing within a sentence. Sarvam and Soniox transcribe callers who mix English into Malayalam; on Soniox, pick the multilingual option. A sentence with a few English words is still treated as Malayalam.
• A full language switch. Add both English and Malayalam to one agent with multilingual_config, each with its own transcriber, voice and prompt. Malayalam is one of the languages Bolna detects automatically, so with the default “requested or auto detected” trigger the agent follows a caller who moves into Malayalam. The multilingual quickstart shows the API shape.

• Start a call in Malayalam with "agent_data": { "language": "ml" } on POST /call, or an agent_language column set to ml in a batch CSV. Inbound calls can return agent_language from your caller data source.
• Let system messages such as the hangup and “are you still there?” checks follow the active language.
• Read How language switching works for what does and doesn’t trigger a switch.

​

Notes

• On Sarvam, use bulbul:v3 for Malayalam. bulbul:v4-flash has no Malayalam voices.
• Voice ids do not carry across Bulbul generations. A bulbul:v4-flash voice fails on bulbul:v3 with 400 Provided voice: <voice> is not available for the provider: sarvam, so pick one of the 16 bulbul:v3 voices.

​

Phone numbers for Malayalam agents in India

• Buy an Indian number in the dashboard through Plivo or Vobiz.
• For commercial or telemarketing calls in India, read How to get 140 & 160-series phone numbers in India for the TRAI and DLT requirements.
• To use numbers from your own carrier, connect it with a SIP trunk.

​

Next steps