Speko releases OpenRouter API to auto-select optimal voice-AI model stacks

Voice-AI deployments typically stitch together three separate models, speech-to-text, a large language model, and text-to-speech, each of which has dozens of vendor options that change monthly. Speko turns the repetitive benchmarking and integration work into an API, allowing a team to request a stack that meets specific accuracy, latency, or cost goals. The service returns the chosen provider names and scores in response headers, automating what was previously a manual, quarterly-or-longer process.

When a request arrives, the router filters its public benchmark database for models that satisfy the requested language, region, and constraints, then selects the top-scoring combination. It prefetches signed session plans so the caller can connect directly to the provider without an additional control-plane round trip, and failover is limited to the connection-setup stage if the primary provider rejects the connection. This design reduces per-call latency and simplifies failover handling for production agents.

Adopting Speko incurs no cost for the open-source gateway binary or for BYOK deployments; pricing applies only to the hosted router service and managed key billing. The gateway runs as a single Go binary sidecar in the agent container, communicating over a Unix socket and optionally pinning provider hosts to avoid sharing keys with Speko’s cloud. Telemetry is enabled by default but can be turned off with a single environment variable, giving operators control over data collection.

The solution is limited to models that Speko has benchmarked and published; users cannot route to custom or unpublished models without adding them to the benchmark set. Failover only occurs during the initial connection attempt, so runtime provider failures after the session starts are not automatically handled. Additionally, the hosted router requires internet access unless the BYOK mode is used, which may be a constraint for isolated environments.

By abstracting model selection into a managed API, teams can continuously benefit from newer, cheaper, or more accurate speech models without redeploying code or re-benchmarking internally. This can improve call transcription accuracy, reduce operating costs, and accelerate feature rollout for voice-driven applications across multiple languages and regions.