Voice Assistant API in India

Build Voice Assistants with Unoblox

Create voice bots for phone calls, IVR, voice commands—combine speech-to-text (STT) with unoblox LLM endpoint, then text-to-speech (TTS) for response. Billed entirely in rupees.

Voice assistant pipeline

1. Listen: Turn speech into text (use Deepgram, AssemblyAI, or open-source Vosk).
2. Understand: Send text to unoblox LLM endpoint.
3. Respond: Turn the reply text into speech (Google TTS, ElevenLabs, or OSS).
4. Act: Execute command (booking, transfer, payment).

Models for voice understanding

Model	Input/Output (₹)	Latency	Use case
Qwen3 Max	₹120.95 / ₹604.77	300–800ms	Complex queries
Claude Sonnet	₹201.6 / ₹1008	200–600ms	Nuanced intent
Qwen3.8-27B	₹16.32 / ₹48.96	100–400ms	Fast responses
Claude Opus	₹504 / ₹2520	400–900ms	Premium reasoning

Build a voice assistant in 3 steps

Step 1: Capture voice and turn it into text

Use a speech-to-text service (Deepgram API call):
POST https://api.deepgram.com/v1/listen?model=nova-2&language=hi-IN
Audio file → "Mere liye ek taxi book kar do"

Step 2: Send text to unoblox LLM

POST https://api.unoblox.ai/v1/messages
Authorization: Bearer ub-gw-...

{
  "model": "qwen/qwen3-max",
  "messages": [{"role": "user", "content": "User request (Hindi): Mere liye ek taxi book kar do. Respond in Hindi, briefly."}],
  "max_tokens": 100
}
Response: "Main aapke liye ek taxi book kar deta hoon. Aapka current location kya hai?"

Step 3: Turn the response into speech + play

Use Google Cloud Text-to-Speech or ElevenLabs:
Text: "Main aapke liye ek taxi book kar deta hoon..."
Output: Audio MP3 → Play on phone/speaker

Real scenario: Hindi voice assistant for taxi booking

1. Customer calls bot; system plays: "Namaste, unoblox taxi ke liye kya chahiye?"
2. Customer says (Hindi): "Sector 10 se sector 20 ek taxi de do."
3. Speech-to-text: Turns the audio into text.
4. LLM call: Extract pickup, dropoff, time.
5. Execute: Book via external API.
6. Respond: "Aapki taxi 5 min mein aayegi. Booking ID: XYZ."
7. Cost: ~100 tokens/call ≈ ₹20–50 on Qwen3 Max.

Multilingual voice support

Language support:

• English: All models (excellent).
• Hindi: Qwen3 Max, Claude Sonnet (good).
• Tamil, Telugu, Kannada: Test on your use case; may need fallback.

Approach:

1. Auto-detect language from STT output.
2. Prompt LLM to respond in same language.
3. Use same language for TTS.

Intent extraction from voice

Ask the LLM to extract structured intent:

Given: "Book me a taxi from Sector 10 to Sector 20, arriving by 3 PM."
Extract JSON: {"intent": "taxi_booking", "pickup": "Sector 10", "dropoff": "Sector 20", "time": "3 PM"}

Frequently asked questions

Q: How do I handle speech recognition errors? A: Use confirmation loops. "I heard: Sector 10 to Sector 20. Is that correct?" Retry if user says no.

Q: Can the voice assistant handle noisy backgrounds? A: STT services (Deepgram, AssemblyAI) are noise-tolerant. Test in your environment; consider noise filtering if needed.

Q: What's the end-to-end latency (speech to speech)? A: ~2–5 seconds (STT ~1s, LLM ~0.5s, TTS ~1s). Acceptable for most use cases; queue requests if volume exceeds capacity.

Q: Can I build this for IVR (phone tree)? A: Yes. Connect to Twilio, Vonage, or open-source FreeSWITCH. Use their webhooks to send audio, receive text, call unoblox, return TTS.

Q: How do I handle sensitive commands (payment, transfer)? A: Use authentication (OTP, voice biometrics) before processing. Always confirm critical actions. Log all transactions.

Q: Does this support Hindi, Marathi, and regional languages? A: Claude Sonnet and Qwen3 Max handle Hindi well. Regional languages are underserved by current models. Recommend testing before production.

Get started in rupees → https://unoblox.ai/sign-in