Boltic: How to Configure Bolti's LLM and Speech Settings

Bolti, a voice AI platform for phone agents, helps you deploy conversational agents that handle outbound sales, support, and scheduling with sub-second latency. Whether you are searching for "Boltic" or looking to optimize your voice setup, understanding the core configurations—specifically how your agent thinks (LLM) and how it hears (Speech)—is key to building a production-ready system. With Bolti's ₹6/min pay-as-you-go pricing and a free trial that includes 50 minutes, you can get a multilingual agent live in minutes.

Two primary tabs in the Bolti dashboard govern the quality of every conversation: the LLM Tab and the Speech Tab. Configuring these correctly prevents lag, improves instruction-following, and ensures your caller's intent is never lost in translation.

What is the Boltic LLM Configuration and How Do You Choose a Model?

The LLM tab controls which large language model generates your agent's replies on every turn of every conversation. For voice-based interactions, choosing the right model requires balancing speed, intelligence, and cost. A single 10-minute call can easily trigger over 60 LLM turns, making model selection highly consequential.

When evaluating models for your voice agent, prioritize these three properties:

• First-Token Latency: Voice agents must respond within ~400ms to feel natural. Smaller, highly optimized models like gpt-5-mini, gpt-5-nano, gemini-2.0-flash-lite, or Groq's llama-3.1-8b-instant excel here.
• Instruction Following: Complex workflows, compliance guardrails, and tool calling require smarter models. If your agent handles sensitive financial transactions or healthcare triage, larger models like gpt-5.1 or gemini-1.5-pro are more reliable.
• Cost per Turn: Because voice calls require frequent, short turns, high-volume deployments benefit from cost-efficient models.

Currently Supported LLM Providers

Bolti supports state-of-the-art models from top providers, allowing you to switch easily depending on your use case:

• OpenAI: gpt-5-mini (default), gpt-5-nano, gpt-5.1, gpt-4o-mini, gpt-4o
• Gemini: gemini-2.0-flash-lite, gemini-2.0-flash, gemini-1.5-pro
• Groq: llama-3.1-8b-instant, llama-3.3-70b-versatile
• DeepSeek: deepseek-chat, deepseek-reasoner
• Baseten: Qwen3-235B-A22B, DeepSeek-V3.1

By default, Bolti uses platform-managed credentials so you can start immediately. If you want to scale cost-effectively or have strict compliance rules, you can bring your own API keys.

How Do You Configure the Speech Tab for Accurate Transcription?

The Speech tab controls how your agent hears by selecting the speech-to-text (STT) engine that transcribes caller audio in real time. Bad transcription is the leading cause of voice agent failures; if the STT engine misinterprets "refund" as "reform," the underlying LLM will provide an irrelevant response.

Bolti supports several enterprise-grade STT providers to ensure high-accuracy transcription:

1. Deepgram: nova-3 (default) and nova-2. This is the recommended choice for almost all use cases due to its low latency and excellent support for regional accents.
2. AssemblyAI: universal-streaming-multilingual. Best for high-accuracy English and multilingual scenarios.
3. ElevenLabs: scribe_v2_realtime. Extremely low-latency real-time transcription.
4. Cartesia: ink-whisper. Supports over 90 languages.
5. Azure: default. Enterprise-grade compliance and integration with your Azure tenancy.

How Do You Handle Indian Accents and Multilingual Calls?

If you are deploying voice agents in India, handling accents, regional languages, and code-switching (mixing Hindi and English) is essential. Setting the correct language code on the Speech tab prevents transcription errors.

We recommend the following configurations for Indian deployments:

• Indian English: Use Deepgram nova-3 with the en-IN language code to accurately capture Indian accents.
• Pure Hindi: Use Deepgram nova-3 with the hi code. If you experience accuracy issues, test Cartesia ink-whisper.
• Hinglish / Code-Switching: If your callers switch between Hindi and English mid-sentence, configure Deepgram nova-3 with the multi setting, or use AssemblyAI's universal-streaming-multilingual model.

To see how businesses apply these configurations in real-world scenarios, explore our Bolti use cases.

How Can You Troubleshoot Agent Behavior and Transcription Issues?

Before changing your LLM or STT provider, check your basic settings. If your agent is going off-topic or ignoring instructions, the issue is usually a loose system prompt. Refine your prompt on the Basic tab before swapping models.

To diagnose transcription errors, review your call logs in the dashboard. The Logs tab displays the exact transcript alongside the recorded audio, allowing you to see if the STT engine is mishearing numbers, names, or specific intents. If you spot consistent errors, try changing the regional language code or switching providers. Check our Bolti pricing page to see how different configurations fit your budget.

Set Up Your First Boltic Voice Agent

Ready to build your own high-performance voice agent? You can configure your LLM, select your STT provider, and have a fully functional conversational agent running in less than 10 minutes.

Sign up for your free Bolti account today to get 50 free minutes of call time, or transition to our ₹6/min pay-as-you-go pricing as you scale up your operations.

Frequently Asked Questions

What is the default LLM for a new Bolti agent?

The default model for new agents is OpenAI's gpt-5-mini. It offers an optimal balance of fast first-token latency, strong instruction-following capabilities, and cost-efficiency for voice-based interactions.

Which STT provider is recommended for Indian English and Hindi?

We recommend Deepgram nova-3. Use the 'en-IN' language code for Indian English, 'hi' for Hindi, and the 'multi' setting for callers who switch between Hindi and English during the conversation.

Can I use my own API keys with Bolti?

Yes. While Bolti provides platform-managed credentials by default, you can bring your own API keys (BYO) for both LLM and STT providers to lower costs at scale or meet compliance requirements.

How do I troubleshoot an agent that is ignoring instructions?

If your agent is misbehaving, the fix is usually on the Basic tab. Tighten your system prompt, refine the conversation goal, and add explicit guardrails before changing your LLM provider.