Speko

Speko is a Y Combinator-backed voice AI routing platform that sits in front of every major speech provider and selects the best model per language and objective from published benchmark measurements. It exposes a single OpenAI-compatible API endpoint, so frameworks like LiveKit and Pipecat can point at Speko without changing agent code. The platform is SOC 2 Type II certified, HIPAA-compliant, and GDPR-ready, and is currently in public preview.

What It Is

Speko is a model router for voice AI pipelines. Instead of picking a single speech-to-text, LLM, or text-to-speech provider and hoping it performs well across all languages, Speko reads its own cross-provider benchmark data and routes each request to the model that scores best for the requested language and objective (latency, quality, cost, or balanced). The router covers all three legs of a voice pipeline — STT, LLM, and TTS — under one API key and one base URL.

How the Routing Works

The core insight behind Speko is that most speech model leaderboards are English-only. According to the homepage, 11 of 23 measured STT models have no published benchmark data outside English, and 4 different models win across just 9 languages. Speko's router uses its own per-language benchmark board to make routing decisions rather than relying on a vendor's English leaderboard.

Routing is controlled through request headers:

• X-Speko-Objective — choose latency, quality, cost, or balanced
• X-Speko-Language — BCP 47 tag for benchmark selection
• X-Speko-Allow / X-Speko-Deny — allowlist or blocklist specific providers or models
• X-Speko-Max-Price — cap the per-unit cost of the selected model

Each response includes X-Route and X-Route-Reason headers so callers can observe which provider was selected and why. Pre-response failover is built in: if a candidate fails, the router tries the next-ranked model before returning an error.

Gateway and Framework Integration

Speko speaks the OpenAI API wire format, so any framework that already calls OpenAI for STT, LLM, or TTS needs only a new base_url and API key. The docs show working examples for:

• LiveKit (TypeScript) — point the OpenAI plugin at https://api.speko.ai/v1 with model: 'auto'
• Pipecat (Python) — same pattern for the pipeline's STT, LLM, and TTS stages
• Claude / Cursor — via MCP at https://mcp.speko.ai/mcp

The auto model string tells the router to rank eligible models dynamically. Callers can also pin a specific provider:model ID from the /v1/models list when they want deterministic routing.

API Surface

The API covers the full voice pipeline with both OpenAI-compatible and Speko-native routes:

• POST /v1/audio/transcriptions and POST /v1/transcribe for batch and streaming STT
• POST /v1/chat/completions for LLM
• POST /v1/audio/speech and POST /v1/synthesize for TTS
• GET /v1/realtime WebSocket for speech-to-speech
• GET /v1/transcribe/stream and GET /v1/synthesize/stream for realtime streaming
• GET /v1/routing/preview to preview a route without sending upstream traffic

The OpenAPI spec is published at https://api.speko.ai/openapi.json.

Current Status

Speko is in public preview with all systems operational. The platform is backed by Y Combinator and offers $100 in signup credit to every new account. Enterprise access is available by contract with committed monthly minimums, dedicated support, and SLA terms. No SLA is offered during the public preview period for the Router and Speko infra tiers.