Skip to content

MANIFESTO.md

Latest commit

 

History

History
258 lines (187 loc) · 6.59 KB

MANIFESTO.md

File metadata and controls

258 lines (187 loc) · 6.59 KB
Raw

The Sovereign Voice Manifesto

Voice AI is broken.

Not because the models are bad —
but because access is rationed by wrappers.

The Problem

Today's "Voice AI platforms" are toll booths.

They:

  • Wrap public APIs (OpenAI, Deepgram, Cartesia)
  • Add minimal orchestration
  • Charge per-minute rent
  • Lock you into black boxes
  • Sit between you and your users' conversations

At scale, this is insane.

You pay:

  • For your LLM
  • For your STT
  • For your TTS
  • For your SIP trunk
  • And again — just to connect them

This is not innovation.
It's rent-seeking.

The Cost Nobody Talks About

Voice agents hear everything:

  • Medical data
  • Financial details
  • Sales conversations
  • Private emotions
  • Business secrets

The Questions You Should Ask

  • Where do your calls go?
  • Who stores the recordings?
  • Who can access the transcripts?
  • What happens to the data after the call?

"Trust us" is not a security model.

When you route calls through a managed platform:

  • You lose data sovereignty
  • You accept their terms
  • You hope they're secured
  • You pay for the privilege

What We Believe

Ownership

  • You should own your calls — not rent access to them
  • You should own your infrastructure — not depend on someone else's
  • You should pay for compute — not permission

Technical Excellence

  • Latency matters — and it's determined by your model choices, not our infrastructure
  • Interruptions matter — Siphon gives you the controls, you tune them
  • Python developers shouldn't be forced into Node or Go to ship voice

Transparency

  • Open source beats black boxes
  • Direct integrations beat wrappers
  • Standards beat vendor lock-in

What Siphon Stands For

Siphon is Sovereign Voice AI.

Principle Implementation
🔓 Open Source Apache 2.0 license — inspect, modify, deploy
🏠 Your Infrastructure Runs on your cloud, your servers, your LiveKit
💰 No Platform Fees Pay your AI providers directly, no markup
🔒 No Hidden Routing Your calls go from SIP → LiveKit → your agent
🔑 No Lock-In Swap LLMs, STT, TTS providers with config changes

Build Agents That Speak For You

Not for a platform.

Siphon is not the easiest path.
It is the right one.

If you want convenience, buy a wrapper.
If you want control, use Siphon.

Technical Recommendations

Note: Latency and interruption handling depend heavily on your model choices. Siphon's infrastructure is optimized for sub-500ms latency — but the models you choose determine the final user experience.

For Low-Latency Production (< 1 second end-to-end)

Best Overall Stack:

LLM: Groq (Llama 3.3 70B) or Cerebras (Llama 3.1 8B)
STT: Deepgram Nova 3
TTS: Cartesia Sonic or ElevenLabs Turbo
  • Why: Groq/Cerebras have 200-500ms time-to-first-token (TTFT)
  • Latency: ~800ms-1.2s total (STT → LLM → TTS → audio)
  • Cost: ~$0.02-$0.05/min (direct provider costs)

For Best Quality (Natural Conversations)

Premium Stack:

LLM: OpenAI GPT-4o or Anthropic Claude 3.5 Sonnet
STT: Deepgram Nova 3 or AssemblyAI
TTS: ElevenLabs (standard) or Cartesia Sonic
  • Why: GPT-4o has excellent conversational ability and function calling
  • Latency: ~1.5-2.5s total
  • Cost: ~$0.05-$0.12/min (direct provider costs)

For Cost-Sensitive Deployments

Budget Stack:

LLM: DeepSeek V3 or Groq (Llama 3.1 8B)
STT: Deepgram Nova 2
TTS: Cartesia Sonic or OpenAI TTS
  • Why: DeepSeek is incredibly cheap, Groq has free tier
  • Latency: ~1-1.5s total
  • Cost: ~$0.01-$0.03/min (direct provider costs)

For Multilingual Support

Global Stack:

LLM: OpenAI GPT-4o or Google Gemini 2.0 Flash
STT: Deepgram Nova 3 (multi-language) or Google STT
TTS: Sarvam AI (Hindi/Indic) or ElevenLabs (multilingual)
  • Why: Gemini 2.0 Flash excels at multilingual, Sarvam optimized for Indic languages
  • Latency: ~1.2-2s total
  • Cost: ~$0.03-$0.08/min (direct provider costs)

For Maximum Context & Complex Tasks

Power Stack:

LLM: Anthropic Claude 3.5 Sonnet (200k context)
STT: Deepgram Nova 3
TTS: ElevenLabs or Cartesia
  • Why: Claude 3.5 handles complex multi-step reasoning
  • Latency: ~2-3s total (worth it for quality)
  • Cost: ~$0.08-$0.15/min (direct provider costs)

Interruption Handling

For natural barge-in, configure:

Agent(
    allow_interruptions=True,
    min_interruption_duration=0.08,  # 80ms minimum to register
    min_endpointing_delay=0.45,      # 450ms before detecting turn end
    activation_threshold=0.4,         # VAD sensitivity (lower = more sensitive)
)

Critical for interruptions:

  • ✅ Fast LLM TTFT — Groq, Cerebras, GPT-4o mini
  • ✅ Streaming TTS — Cartesia, ElevenLabs Turbo
  • ✅ Low endpointing delay — 400-600ms sweet spot

Comparison: Managed Platform vs Siphon

Aspect Managed Platform (Vapi/Retell/Bland) Siphon (Self-Hosted)
Cost (10k minutes) $500-$3,100 + AI costs $0 + AI costs ($200-$800)
Data Routing Platform servers → your users Your servers → your users
Vendor Lock-In Hard (API-specific) None (swap providers anytime)
Observability Platform dashboard Full control (your logs, your storage)
Customization Limited to API parameters Full code access
Compliance Trust their security You control security
Latency Platform routing overhead Direct LiveKit connection
Scaling Platform limits Your infrastructure limits

The Choice

You have two paths:

Path 1: Managed Platform

  • ✅ Quick to start
  • ✅ Managed infrastructure
  • ❌ Per-minute fees ($0.05-$0.30/min)
  • ❌ Data goes through their servers
  • ❌ Limited customization
  • ❌ Vendor lock-in

Path 2: Siphon (Sovereign)

  • ✅ No platform fees
  • ✅ Full data ownership
  • ✅ Complete customization
  • ✅ No vendor lock-in
  • ❌ Requires setup (LiveKit + SIP)
  • ❌ You manage infrastructure

Choose wisely.

Join the Movement

Siphon is more than a framework.
It's a statement.

The future of voice AI is sovereign.

Built with Siphon.
Owned by you.