Skip to main content
⚠️ This website is a personal portfolio showcasing my technical capabilities. SabaTech is a brand name, not a legal entity.
Voice POC — v2.3.1

Conversational Voice Assistant

100% Offline AI Voice Pipeline

A complete real-time voice assistant running entirely on your own GPU — ASR → LLM → TTS with sentence-by-sentence streaming and instant barge-in.

Qwen3-ASR transcribes, llama.cpp runs Gemma 4 12B QAT with MTP speculative decoding (148 tok/s), Supertonic 3 synthesises speech, and a WebSocket streams audio back — all without a single packet leaving your machine.

End-to-End Pipeline

From Microphone to Voice, On One GPU

Five stages, zero cloud. Each block runs locally on an NVIDIA RTX 4080 (16 GB VRAM) and chains into the next over a WebSocket.

🎤
1

Stage 1

Qwen3-ASR 0.6B

asistente.pipe.asr.desc

🧠
2

Stage 2

Gemma 4 QAT + MTP

asistente.pipe.llm.desc

🔊
3

Stage 3

Supertonic 3 ONNX

asistente.pipe.tts.desc

4

Stage 4

WebSocket binary

asistente.pipe.streaming.desc

5

Stage 5

Silero VAD

asistente.pipe.bargein.desc

<2s

E2E Latency (MTP)

~9 GB

VRAM / Session

5

TTS Engines

100%

Offline

What Makes It Different

Not a wrapper around a cloud API — a full voice stack that runs on a single GPU

🔒

Total Privacy

100% offline. Audio, transcripts and models never leave your hardware. No vendor, no leak surface, full European data sovereignty.

🔁

Complete Pipeline

ASR → LLM → TTS → streaming → barge-in, end-to-end in one codebase. No gluing five SaaS together — every stage is yours.

🎙️

Multi-TTS

Five interchangeable engines: Supertonic 3 (production), Kokoro 82M (lowest latency, RTF 0.02), edge-tts (online), Piper and Qwen3-TTS (offline, high quality).

Real-time Barge-in

Neural Silero VAD detects the user speaking over the assistant and interrupts playback instantly — natural turn-taking, not walkie-talkie.

📋

Integrated CRM & Signing

Glassmorphism frontend with live chat transcript, customer CRM panel and AES digital signature via Signaturit (SMS OTP).

MTP Speculation

Gemma 4 QAT + a 242 MB MTP drafter verify tokens speculatively — 148 tok/s vs 84 tok/s without MTP, a 75% speedup.

Pricing

Open core for non-commercial use, commercial licence for production deployments

PolyForm Noncommercial

Core

Research, education & hobby

Free
  • Full ASR → LLM → TTS pipeline
  • 5 swappable TTS engines
  • WebSocket streaming + barge-in
  • Frontend with CRM & digital signature
  • Complete source code
Get Core
Best Value

Enterprise

Multi-tenant outbound

Custom
  • Everything in Core
  • Native multi-tenant (PostgreSQL + Redis)
  • Concurrent outbound calls (10+)
  • SIP trunking / real telephony
  • Dual legal compliance
  • On-premise deployment
Talk to Sales

Custom

Tailored deployment

Custom
  • Everything in Enterprise
  • Custom CRM / dialer integrations
  • Fine-tuned models (LoRA / QLoRA)
  • Additional languages & accents
  • 24/7 SLA support
Contact Sales

Want a voice assistant that never phones home?

Self-hosted. Offline. European. Clone the PolyForm Noncommercial core, or talk to us about a multi-tenant production deployment.