# Speko

> A voice AI router that benchmarks speech models language by language and routes STT, LLM, and TTS calls to the best-performing provider through a single OpenAI-compatible API.

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.

## Features
- Cross-provider voice AI routing (STT, LLM, TTS)
- Per-language benchmark-driven model selection
- OpenAI-compatible API (drop-in base URL swap)
- Pre-response failover across providers
- Objective-based routing: latency, quality, cost, balanced
- Language-aware routing via BCP 47 tags
- Provider allowlist and blocklist controls
- Max-price constraint per request
- X-Route and X-Route-Reason response headers for observability
- Realtime speech-to-speech WebSocket
- Realtime STT and TTS streaming WebSockets
- Route preview without upstream traffic
- LiveKit and Pipecat integration
- MCP server for Claude and Cursor
- SOC 2 Type II, HIPAA, and GDPR compliance
- $100 signup credit for all accounts

## Integrations
LiveKit, Pipecat, OpenAI, Deepgram, AssemblyAI, ElevenLabs, Cartesia, Google Cloud Speech (Chirp 3), xAI (Grok STT), Gladia, Smallest AI, Soniox, Modulate, Inworld, Alibaba (Qwen3-ASR), Gradium, Claude (MCP), Cursor (MCP)

## Platforms
WEB, API, CLI

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://speko.ai
- Documentation: https://speko.ai/docs
- EveryDev.ai: https://www.everydev.ai/tools/speko