> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gosutox.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Voice Models & Audio Latency

> Neural voice model selection, sub-300ms duplex streaming WebSocket audio pipeline, and latency optimization.

## Real-Time Duplex Audio Architecture

Traditional voice interfaces rely on sequential HTTP requests: record audio, send to server, convert to text, query LLM, generate speech, and download an audio file. This serial cascade results in 3 to 8 seconds of latency, destroying natural conversational rhythm.

**AXON Talk** solves this with a **Pure Duplex Streaming WebSocket Architecture**:

```mermaid theme={null}
flowchart LR
    Mic["🎙️ Raw Audio Input (PCM / Opus)"] --> WS_In["⚡ Streaming WebSocket Ingress"]
    WS_In --> Fast_ASR["🧠 Real-Time Neural Speech-to-Text"]
    Fast_ASR --> LLM_Stream["⚡ Streaming Brain Reasoner"]
    LLM_Stream --> Neural_TTS["🔊 Ultra-Fast Neural Voice Synthesis"]
    Neural_TTS --> Audio_Out["🎧 Streaming Audio Playback (< 300ms)"]
```

***

## Technical Performance Specifications

| Performance Metric                 | AXON Talk Standard                           | Legacy Voice Bots                     |
| :--------------------------------- | :------------------------------------------- | :------------------------------------ |
| **Acoustic Round-Trip Turnaround** | **\< 300ms**                                 | 3,000ms – 8,000ms                     |
| **Transport Protocol**             | Bi-directional streaming WebSockets          | Sequential HTTP REST requests         |
| **Interruption Handling**          | Instant voice barge-in (\< 50ms cut-off)     | Cannot interrupt until audio finishes |
| **Audio Fidelity**                 | 24kHz studio-grade neural voice synthesis    | 16kHz robotic text-to-speech          |
| **Multilingual Parity**            | Native English, Korean, and Chinese phonemes | Heavily accented translated phonemes  |

***

## Neural Voice Model Selection

AXON Talk provides curated neural voice profiles engineered for professional executive communication:

<CardGroup cols={3}>
  <Card title="Executive Authority (Male / Female)" icon="crown">
    **Confident, Measured & Clear**
    Calibrated for boardroom presentations, legal cross-examination sparring, and strategic advisory.
  </Card>

  <Card title="Warm & Socratic (Male / Female)" icon="heart">
    **Engaging, Patient & Expressive**
    Optimized for executive coaching, language training, and educational mentorship.
  </Card>

  <Card title="High-Velocity Analyst" icon="gauge-high">
    **Crisp, Efficient & Direct**
    Tuned for rapid morning market news summaries and dense financial data briefings.
  </Card>
</CardGroup>

***

## Voice Barge-In & Acoustic Echo Cancellation

* **Instant Voice Barge-In**: When you speak while the AI is talking, AXON Talk detects your voice within 50ms, immediately halts downstream audio playback, and seamlessly shifts back to listening mode without turn desynchronization.
* **Acoustic Echo Cancellation (AEC)**: Filters out speaker audio bleed so you can converse naturally using laptop speakers without requiring headphones.
* **Adaptive Jitter Buffering**: Dynamically compensates for fluctuating mobile cellular networks (4G/5G/Wi-Fi) to eliminate audio dropouts while driving or walking.

***

## Next Steps

<CardGroup cols={2}>
  <Card title="Mobile & Hands-Free Scenarios" icon="car-side" href="/talk/mobile-handsfree-scenarios">
    Explore in-car briefings, language training, and walking brainstorming.
  </Card>

  <Card title="Live Data Architecture" icon="database" href="/data/live-data-architecture">
    Learn how real-time market APIs and multimodal vault assets feed live data into AXON Talk.
  </Card>
</CardGroup>