Speko – OpenRouter for Voice AI

TL;DR

TL;DR: Speko is a voice AI router that benchmarks STT, LLM, and TTS models across 10 languages and exposes them through a single OpenAI-compatible API — letting you swap providers without touching your framework code.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: speko.ai — verified via direct fetch
• HN launch thread: news.ycombinator.com/item?id=42310412 — confirmed via HN Algolia API
• YC company page: ycombinator.com/companies/speko — confirmed via product footer link
• License: Proprietary (SaaS, no open-source repository)
• Source last checked: 2026-08-19

What Is Speko?

Voice AI pipelines typically chain three model types: speech-to-text (STT), a language model (LLM) for reasoning, and text-to-speech (TTS). Each has its own provider, pricing model, and API surface. Switching from Deepgram to ElevenLabs for TTS means touching real code.

Speko’s pitch is that this glue logic — routing, benchmarking, failover — should be someone else’s problem. It benchmarks models in-house across real audio in 10 languages (English, Spanish, Mandarin, Arabic, Hindi, Portuguese, French, German, Japanese, Korean) and exposes a single gateway that speaks the OpenAI Realtime API.

From the product page:

“Speko speaks the OpenAI API, so the frameworks you already use need a hostname and a model string.”

That means LiveKit, Pipecat, and other frameworks that already support the OpenAI Realtime API can point at Speko’s gateway without code changes.

How the Routing Works

Speko publishes benchmark data at benchmarks.speko.ai. The benchmarks cover:

• STT: Word Error Rate (WER), hallucination rate, dead-air percentage, stall percentage
• LLM: fabrication rate, time-to-first-token, dead-air percentage, tool silence percentage
• TTS: synthesis latency, character cost per million, MOS/ELO scores

For a given task, Speko maintains a ranked leaderboard. Based on the measured metrics, it routes each session to the best-performing available model — or lets you pin a specific stack for predictable latency.

Example stack from their benchmarks page:

Real-time phone agent:
  STT   → Universal-3.5 Pro
  LLM   → gemma-4-31b
  TTS   → inworld-tts-2
  Latency: 66ms · 192ms · 116ms

The stacks page shows full breakdown by use case: real-time phone agent, accuracy-critical, natural conversation, tool-heavy agent.

API Overview

The gateway exposes three model categories:

| Category | Example model ID | Notes | |---|---|---| | stt | deepgram:universal-3.5-pro | In-house benchmarks ranked | | llm | anthropic:claude-haiku-4.5 | Via OpenAI-compatible chat completions | | tts | cartesia:sonic-3.5 | Realtime TTS via OpenAI-compatible endpoint |

Pricing is pay-as-you-go at $0.09 per minute for routed calls, with a free tier available at sign-up. The gateway handles provider failover automatically.

Setup

1. Create an account at platform.speko.ai
2. Get an API key from the dashboard
3. Point your framework at the Speko gateway URL with your model string

No SDK install needed if your framework already speaks the OpenAI Realtime API — just swap the base URL and model ID.

Supported Frameworks

Based on docs links visible on the site:

• LiveKit
• Pipecat

Both have dedicated integration pages under /docs/.

Security and Compliance

The footer lists SOC 2 Type II, HIPAA, and GDPR badges — relevant if you handle voice data in regulated industries.

Practical Evaluation Checklist

• Free tier available for prototyping
• OpenAI-compatible API means no new SDK for most frameworks
• Benchmarks are provider-side (not third-party) — directly informs routing decisions
• Multi-language leaderboard covers 10 languages
• SOC 2 / HIPAA / GDPR badges for enterprise evaluation
• No GitHub repo (proprietary SaaS)

FAQ

Q: Does Speko host the models or just route to them? A: It routes. Speko sits in front of existing providers (Deepgram, ElevenLabs, Anthropic, etc.) and routes traffic based on its own benchmark data.

Q: How does Speko choose which model to use? A: It maintains a ranked leaderboard per task type and language, updated as it measures real traffic. You can use the automatic router or specify a pinned stack.

Q: What happens if a provider goes down? A: The gateway handles failover across providers. Specific SLA terms would be in the enterprise contract.

Q: Is there a self-hosted option? A: No. Speko is a hosted service only.

Q: Does it support real-time voice calls (WebRTC)? A: Yes. The LiveKit and Pipecat integrations suggest WebRTC-based pipelines are supported.

Conclusion

Speko solves the voice AI integration tax: the cognitive load of evaluating and swapping between STT, LLM, and TTS providers across languages. By publishing in-house benchmarks and wrapping everything in an OpenAI-compatible gateway, it lets teams focus on product logic rather than provider glue.

The pricing model ($0.09/min, free tier) is approachable for prototyping, and the compliance badges (SOC 2, HIPAA, GDPR) suggest it’s being positioned seriously for production voice agents in healthcare, customer service, and other regulated domains.