> For the complete documentation index, see [llms.txt](https://dataqueue.gitbook.io/voicehub-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://dataqueue.gitbook.io/voicehub-docs/getting-started/how-voicehub-works.md).

# How VoiceHub Works

### **Core Models**

VoiceHub operates as an orchestration layer over three core components in the voice AI pipeline:

1. Transcriber (STT) — Converts audio input into text
2. Language Model (LLM) — Interprets the text and generates intelligent responses
3. Voice (TTS) — Converts the LLM’s response back into speech

These modules can be flexibly configured using top-tier providers:

* STT: Deepgram, Gladia, Azure, etc.
* LLM: OpenAI, Groq, Claude, Cohere, etc.
* TTS: ElevenLabs, PlayHT, LMNT, etc.

VoiceHub handles orchestration, streaming, and optimization across the three components — ensuring real-time interaction, smooth latency, and seamless switching between providers.

***

### **The Voice-to-Voice Pipeline (Real-Time Streaming)**

**Step 1: Listen (Intake Raw Audio)**

User speaks into their device (laptop, phone, etc.). Audio is streamed and recorded in real time. That audio is then transcribed by the selected STT engine into text.

**Step 2: Understand (Run an LLM)**

The transcribed text is sent to the selected LLM model. That model uses the agent’s prompt and context to generate a response.

**Step 3: Speak (Text → Raw Audio)**

The response text is passed to a TTS engine, which synthesizes speech audio and streams it back to the user.

🎯 All three steps are optimized for real-time execution, targeting end-to-end latency of <200–500ms, depending on configuration.

***

### **What Makes VoiceHub Unique**

* You can switch providers at each stage without writing custom glue code
* We stream audio and text between stages for responsiveness down to \~200ms
* Our routing system handles scaling, retrying, failover, and QoS behind the scenes

***

### **Built-in DQ Models**

For teams focused on Arabic, English or Dutch support, VoiceHub offers in-house DataQueue models:

* Optimized for MENA voice patterns
* Fine-tuned for dialect-specific recognition and tone
* Lower latency and higher reliability than many generic models

Use DQ Mode for:

* Arabic-first customer service
* Multilingual deployments with low setup overhead
* Government, telco, or regulated deployments with regional preferences

You can switch to DQ Mode at any time from the Configuration panel.