Sovereign AI

Run your entire voice AI stack on Australian infrastructure. LYRE integrates with SCX.ai to provide sovereign LLM inference and STT, hosted in NEXTDC data centres in Sydney. No data leaves Australia.

Why sovereign AI?

For organisations in government, healthcare, finance, and legal, data sovereignty is not optional. Sending voice data to overseas providers creates regulatory risk under APP 8 of the Privacy Act and sector-specific rules like APRA CPS 234.

•Data residency guarantee: All audio, transcripts, model weights, and inference happen within NEXTDC Sydney (S1/S2/S3). Nothing crosses Australian borders.

•APP 8 compliance: By using sovereign providers, you eliminate the need for cross-border disclosure assessments entirely.

•No foreign access: SCX.ai infrastructure is Australian-owned and operated. Not subject to US CLOUD Act, FISA 702, or similar foreign data access laws.

•Low latency: Sydney-hosted inference means sub-500ms round-trip times for Australian users, compared to 200-400ms of added latency for US-based providers.

SCX.ai overview

SCX.ai provides GPU inference infrastructure hosted in NEXTDC data centres across Australia. LYRE uses SCX as a sovereign provider for both LLM and STT workloads.

Detail	Value
Base URL	https://api.scx.ai/v1
API compatibility	OpenAI-compatible (chat completions, embeddings, transcriptions)
Authentication	Bearer token (SCX API key)
Data centre	NEXTDC S1/S2/S3, Sydney, Australia
GPU hardware	NVIDIA H100 and A100 clusters
Certifications	SOC 2 Type II, ISO 27001, IRAP (in progress)

Sovereign LLM models

SCX.ai hosts a range of open-weight models optimised for Australian voice agent workloads.

Model	Parameters	TTFT	Tool calling	Best for
MAGPiE 117B	117B	~400ms	Yes	Highest quality sovereign model. Complex conversations.
gpt-oss-120b	120B	~350ms	Yes	Primary recommended model. Good speed/quality balance.
DeepSeek R1	671B (MoE)	~600ms	Yes	Reasoning tasks. Higher latency but strongest analytical ability.
Llama 3.3 70B	70B	~250ms	Yes	Fast general-purpose. Same model as Groq but sovereign.
Qwen 2.5 72B	72B	~280ms	Yes	Strong multilingual. Good for diverse caller populations.

Sovereign STT (Speech-to-Text)

SCX.ai hosts Whisper Large v3 for sovereign speech-to-text. This runs on the same NEXTDC infrastructure, ensuring voice audio never leaves Australia.

Model	Mode	Languages	Notes
whisper-large-v3	Batch	99+	Best for post-call transcription. Not real-time streaming.
whisper-large-v3-turbo	Batch (fast)	99+	Distilled model. Faster at slight quality cost.

Note: SCX Whisper is batch-mode only. For real-time streaming STT with sovereign requirements, pair SCX Whisper (post-call) with Deepgram or Google Cloud STT (Sydney region) for live transcription.

Configuration

Register SCX as an LLM provider

curl

curl -X POST https://api.lyre.au/v1/providers \
  -H "Authorization: Bearer $LYRE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "llm",
    "provider": "scx",
    "api_key": "scx_...",
    "base_url": "https://api.scx.ai/v1",
    "models": [
      "gpt-oss-120b",
      "magpie-117b",
      "llama-3.3-70b",
      "deepseek-r1",
      "qwen-2.5-72b"
    ]
  }'

Register SCX as an STT provider

curl

curl -X POST https://api.lyre.au/v1/providers \
  -H "Authorization: Bearer $LYRE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "stt",
    "provider": "scx",
    "api_key": "scx_...",
    "base_url": "https://api.scx.ai/v1",
    "models": ["whisper-large-v3", "whisper-large-v3-turbo"]
  }'

Create a sovereign agent

curl — fully sovereign agent

curl -X POST https://api.lyre.au/v1/agents \
  -H "Authorization: Bearer $LYRE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Sovereign Agent — Gov",
    "first_message": "Hello, this is the Department of Services AI assistant. How can I help you today?",
    "system_prompt": "You are a helpful government services assistant...",
    "stt": {
      "provider": "deepgram",
      "model": "nova-2",
      "language": "en-AU"
    },
    "llm": {
      "provider": "scx",
      "model": "gpt-oss-120b",
      "temperature": 0.5,
      "fallback": {
        "provider": "scx",
        "model": "llama-3.3-70b"
      }
    },
    "tts": {
      "provider": "google",
      "model": "neural2",
      "voice_id": "en-AU-Neural2-A"
    },
    "compliance": {
      "ai_disclosure": true,
      "recording_consent": "all_party",
      "dncr_check": true
    }
  }'

This configuration keeps LLM inference sovereign via SCX. For maximum sovereignty, pair with Google Cloud STT (Sydney) and Google Cloud TTS (Sydney) to keep the entire pipeline within Australia.

Direct SCX.ai API usage

SCX.ai exposes an OpenAI-compatible API. You can test models directly before configuring them in LYRE.

curl — chat completion via SCX

curl https://api.scx.ai/v1/chat/completions \
  -H "Authorization: Bearer scx_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-oss-120b",
    "messages": [
      {"role": "system", "content": "You are a helpful Australian assistant."},
      {"role": "user", "content": "What are the ACMA calling hour restrictions?"}
    ],
    "temperature": 0.7,
    "max_tokens": 512,
    "stream": true
  }'

Python — using OpenAI SDK with SCX

from openai import OpenAI

client = OpenAI(
    api_key="scx_...",
    base_url="https://api.scx.ai/v1",
)

response = client.chat.completions.create(
    model="gpt-oss-120b",
    messages=[
        {"role": "system", "content": "You are a helpful Australian assistant."},
        {"role": "user", "content": "Explain APP 8 of the Privacy Act."},
    ],
    temperature=0.7,
    stream=True,
)

for chunk in response:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

Data residency architecture

When using the sovereign configuration, the complete data flow stays within Australia:

Sovereign data flow

User (Australia)
  │
  ├── WebRTC ──> LiveKit (Vultr Sydney)
  │                │
  │                ├── Audio ──> Deepgram / Google STT (Sydney region)
  │                │               │
  │                │               └── Text ──> SCX.ai LLM (NEXTDC Sydney)
  │                │                              │
  │                │                              └── Response ──> Google TTS (Sydney)
  │                │                                                 │
  │                └── Audio <──────────────────────────────────────────┘
  │
  ├── Recording ──> Vultr Object Storage (syd1)
  └── Transcript ──> Vultr Postgres (Sydney)

All components: Australian data centres
No cross-border data transfer

When to use sovereign providers

Use case	Recommended config	Reason
Government services	SCX LLM + Google Sydney STT/TTS	IRAP compliance, no foreign data access
Healthcare / telehealth	SCX LLM + Deepgram STT	Patient data must stay onshore (My Health Records Act)
Financial services	SCX LLM + Google Sydney STT/TTS	APRA CPS 234 outsourcing requirements
Legal	SCX LLM + SCX Whisper	Legal professional privilege, client confidentiality
General commercial	Groq LLM + Deepgram STT + ElevenLabs TTS	Lowest latency. Sovereign not required unless processing sensitive data.

Pricing considerations

Sovereign inference on SCX.ai is priced at a premium compared to US-based providers, reflecting the cost of Australian GPU infrastructure. Contact support@lyre.au for volume pricing.

For many organisations, the compliance risk reduction and avoided legal costs of data sovereignty significantly outweigh the infrastructure premium. LYRE also supports hybrid configurations where sensitive calls use sovereign providers and general calls use standard providers.