For teaching, learning, or anything you can dream up.
In any language — or more than one.
Under the hood
Every spoken word travels through seven stages before a frame is rendered. Total round-trip latency is under 1 second.
Mic input
mobile / web
Silero VAD
end-of-speech
Whisper STT
transcription
GPT-4o-mini
response
LoqalKit TTS
synthesis
Voqal GPU
lipsync render
LiveKit stream
WebRTC output
~840ms
total round-trip latency
1.32s
real-time budget per chunk
5.5× headroom at full load
Mic → VAD → Whisper → GPT-4o-mini → TTS → Voqal GPU → LiveKit
Capabilities
Voqal-1.3B processes 33-frame chunks in approximately 170ms. Each inference chunk runs in a serialised FIFO queue shared across sessions, with no per-session GPU overhead.
Arabic and Hebrew ship as first-class citizens with right-to-left text shaping. LoqalKit TTS handles phonetic rendering natively, with no transliteration hacks.
LoqalKit TTS converts LLM output to 16kHz PCM before it reaches Voqal, keeping the voice pipeline modular. Swap TTS providers without touching the rendering stage.
End-to-end round-trip from user speech to rendered video frame is approximately 840ms. The 1.32s real-time frame budget provides 5.5x headroom, plenty of margin for network jitter.
Each session runs in complete isolation. When it ends, all resources are fully reclaimed with no state bleeding between users and no memory accumulation over time.
Voqal publishes rendered video and audio directly to a LiveKit room. It integrates with any WebRTC infrastructure and requires no additional tooling beyond a session token.
Comparison
Side-by-side with the leading cloud avatar platforms.
| Feature | Voqal by Loqal | Tavus | Simli | HeyGen |
|---|---|---|---|---|
| Output FPS | 25fps | ~24fps | 25fps | ~25fps |
| Response latency | <1s | 2–5s | ~1–2s | 3–8s |
| RTL language support | ✓ | ✗ | ✗ | partial |
| Concurrent sessions | 7×/worker | cloud managed | cloud managed | cloud managed |
| Pricing model | usage-based | per minute | per minute | per minute |
| Model size | 1.3B | proprietary | proprietary | proprietary |
Latency and FPS figures are approximate and sourced from public documentation and community benchmarks (2024–2025).
Use cases
Language learning
Voqal powers spoken conversation sessions where learners practice with an AI avatar that hears, thinks, and responds in their target language, complete with natural lipsync.
Enterprise L&D
Generate presenter-led training content in any of 150+ languages without re-shooting. Voqal renders realistic avatar video from text in seconds. Localise once, deploy everywhere.
Live streaming and events
Run a single live AI presenter that speaks to different audience segments in their native language simultaneously. Voqal's WebRTC output drops into any streaming setup.
FAQ
Get started
Voqal runs on Loqal infrastructure. Drop it into any WebRTC or LiveKit setup and your users are talking to a live avatar in minutes.
