Last updated: 2026-11-02
Speko Router is a hosted API that sends speech-to-text, LLM and text-to-speech requests to one of 106 catalog models, chosen by latency, quality or cost. Its benchmarks board covers 17 languages, and every response reports the provider and model that served it. The open Gateway is MIT licensed.
About Speko
Speko is a router for voice AI. Speko Group, Inc., a US company in Y Combinator's Summer 2026 batch, runs a hosted API that accepts speech-to-text, LLM and text-to-speech requests and sends each one to a provider it selects, the same job OpenRouter does for text models. The pitch is that voice teams should not wire a transcription vendor, a language model and a speech vendor together on their own.
The hosted Router exposes typed routes for model discovery, transcription, speech synthesis and LLM responses, plus WebSocket routes for streaming and realtime speech. In auto mode it chooses among eligible routes on speed, accuracy, price or a mix of the three, reading from a public benchmarks board that scores models language by language. In pinned mode you name the provider and model. If a provider fails before output starts, Router can fall back to another eligible route, and every response reports which route served it. Providers on the board include Deepgram, AssemblyAI, ElevenLabs, Fish Audio and Cartesia. The board also lists speech models from OpenAI, xAI and Alibaba Cloud.
Two other products sit beside it. The open Gateway, published under an MIT licence, runs inside your own LiveKit or Pipecat worker and keeps bring-your-own provider keys local. The hosted Agents API handles agents, sessions, phone numbers and calls, which puts it near Vapi and Retell AI for teams that want a managed conversation loop. The choice between the three comes down to who should hold the credentials and own the call.
Developers integrate through public OpenAPI and AsyncAPI documents rather than an OpenAI-compatible hostname, which is a real adjustment if your code already targets the OpenAI SDK. Teams that want a self-hosted text-model proxy are better served by LiteLLM. An unauthenticated docs MCP server and an official CLI help coding agents set it up. The service is still in preview, with email support and no uptime guarantee, so production buyers should confirm terms with the team first. To compare it against other voice options, try Hokai Smart Match.
Pricing
Router: the selected provider's published rate plus 5%. Speko infrastructure: $0.09 per minute covering speech-to-text, LLM and text-to-speech. Enterprise: custom, with a committed monthly minimum.
$10 signup credit on every account. Phone calls on Speko numbers add the carrier charge plus 20%, at least $0.01 per minute inbound and $0.015 outbound. Read from speko.ai/pricing.md on 2026-11-02.
Feature Comparison by Tier
| Feature | Router | Enterprise | Speko infrastructure |
|---|---|---|---|
| Model usage (speech-to-text, LLM, TTS) | Billed at the provider rate | Included | Included in the per-minute rate |
| Routing | Measured selection with failover before output | Adds deployment reviews | Measured selection with failover before output |
| Access | Public preview | By contract | Public preview |
| Support | Dedicated | ||
| Uptime guarantee | None during preview | Written into the contract | None during preview |
| Commitment | None | Committed monthly minimum | None |
Key Features
- Model routing: Auto mode picks a route by latency, quality, cost or a balanced objective, or you pin a provider and model, across a catalog of 106 models spanning speech-to-text, LLM, text-to-speech and speech-to-speech.
- Pre-response failover: If a provider errors before output starts, Router can retry on another eligible provider, and response headers name the provider, model, region and attempt that produced the answer.
- Language-by-language benchmarks: A public board scores speech models on FLEURS audio per language and reports latency percentiles measured from US East, so routing can lean on measured results instead of vendor claims.
- Open Gateway for LiveKit and Pipecat: An MIT-licensed runtime keeps bring-your-own provider keys inside your own process and plugs into LiveKit and Pipecat sessions as STT, LLM and TTS services.
- Typed API contracts: Public OpenAPI 3.1 and AsyncAPI 2.6 documents, a required Idempotency-Key on every POST and one error envelope with a stable code and a retryable flag make retries safe to automate.
- Agent-ready access: A docs MCP server needs no login, an authenticated MCP server exposes account actions, and an official CLI can place disclosed phone calls.
Pros
- Every response names the provider, model, region and attempt that served it, so a bad transcript can be traced to one concrete route.
- The Gateway is MIT licensed and keeps bring-your-own provider keys in the customer's own process, which suits teams that cannot hand keys to a third party.
- Public OpenAPI and AsyncAPI contracts plus an unauthenticated docs MCP let coding agents integrate without guessing at compatibility.
Cons
- Router is not an OpenAI-compatible endpoint, so code written for the OpenAI SDK has to be rewritten against Speko's own contracts.
- Access is a public preview with no SLA and email-only support, which rules it out for workloads that need contractual uptime today.
- The company is young: the Gateway repository was created in August 2026 and the team is a Summer 2026 YC batch, so the track record is short.
- The models page says live availability could not be verified against its saved catalog, so a listed model may not be routable on a given day.
Product Information
- Cloud
- Yes
- Training
- Documentation, Changelog feed, Public benchmarks board
Data Handling
- Training-data policy
- The privacy policy does not state a training policy for voice and API content; it says Google user data is not used to train generalized models.
- Data retention
- 30 days
Frequently Asked Questions
What does Speko cost per minute?
Routed through Router, the bill is whichever provider rate gets picked plus a 5% fee. Speko infrastructure is a flat $0.09 per minute across all three legs, and Enterprise terms are negotiated. New accounts start with $10 in credit, and carrier fees for phone numbers come on top.
Can I try Speko without paying?
There is no permanent free plan. Each new account gets signup credit that is applied before any paid usage, and calls that ring out, go unanswered or fail on Speko's side are not billed. The preview carries no SLA, so use the trial to test routing quality rather than uptime.
Which tools should I weigh against Speko?
For text models behind one API, OpenRouter is the closest model, and LiteLLM suits teams that want to self-host the proxy. For a managed voice-agent stack with phone numbers, Vapi and Retell AI cover that ground. Going direct to Deepgram or ElevenLabs is simpler if you need one vendor and never plan to switch.
How is Speko different from OpenRouter?
OpenRouter routes text models. Speko routes speech-to-text, LLM and text-to-speech together, picks models using language-by-language audio benchmarks, and is not OpenAI-compatible, so you code against its own OpenAPI and AsyncAPI contracts. Pick Speko when audio quality and latency per language drive the decision.
How do I make a first request to Speko?
Create an account on the platform, generate an API key and export it as SPEKO_API_KEY. Call GET /v1/models to list routable models, then POST to /v1/stt/transcriptions, /v1/tts/speech or /v1/llm/responses with an Idempotency-Key header. LiveKit and Pipecat users can use the native Gateway integrations instead.
Top Alternatives
- OpenRouter: Pick OpenRouter if your workload is text models only and you want an OpenAI-style endpoint.
- Retell AI: Choose Retell AI when you want a finished phone-agent product instead of routing models yourself.
- LiteLLM: Choose LiteLLM when the proxy must run in your own infrastructure and speech is not in scope.