Speko

Product Introduction

1. Definition: Speko is a managed API routing platform and data plane specifically engineered for multimodal voice AI workloads. It operates as a technical abstraction layer between applications and a curated marketplace of third-party speech-to-text (STT), large language model (LLM), and text-to-speech (TTS) providers.
2. Core Value Proposition: Speko exists to eliminate the complexity, vendor lock-in, and performance uncertainty in building production-grade voice AI applications. Its primary value is delivering optimal accuracy and cost-efficiency per language through intelligent, benchmark-driven routing, coupled with a flexible architecture for both managed and self-managed deployments.

Main Features

1. The Hosted Router: This is Speko's core managed service at router.speko.dev. It provides a unified, typed API contract for STT, LLM, and TTS calls. The system automatically routes each request to the best-performing provider for the given task, language, and objective (e.g., "balanced," "cost-optimized"). It handles API key management, load balancing, fallback logic, and observability, acting as a fully managed data plane.
2. Public, Language-Specific Benchmarks: Speko maintains and publishes continuous, independent benchmarks of speech AI models across multiple languages. This isn't just high-level performance data; it includes granular Word Error Rate (WER) scores and cost-per-minute metrics, allowing for direct comparison. The benchmarks reveal critical insights, such as which model leads for Tamil versus German, informing the Router's intelligent routing decisions.
3. The Open-Source Gateway: This is a customer-side runtime designed for frameworks like LiveKit and Pipecat. It enables provider-direct streaming where audio streams connect directly from the user's infrastructure to the chosen AI provider (e.g., OpenAI, Deepgram), using the customer's own API keys (BYOK). The Gateway can also leverage Speko's benchmark data to make local routing decisions or proxy calls through the managed Router, offering a hybrid architecture for maximum control and flexibility.

Problems Solved

1. Pain Point: The "best" speech AI model is not universal; it varies drastically by language, accent, and use case. Developers face significant research overhead, integration complexity, and risk of vendor lock-in when trying to manually test and integrate multiple providers to achieve global coverage and optimal performance.
2. Target Audience: Voice AI Application Developers building real-time conversational agents, call centers, or voice interfaces; Product Teams needing reliable, cost-effective multilingual STT/TTS; DevOps & Platform Engineers responsible for the scalability, observability, and resilience of voice AI pipelines.
3. Use Cases: A global customer service bot that automatically uses the most accurate STT model for a caller's language; a content transcription service that dynamically routes files to the most cost-effective provider based on audio quality and required turnaround time; a live video platform integrating real-time captions using the Gateway for low-latency, direct provider connections.

Unique Advantages

1. Differentiation: Unlike single-provider SDKs or generic API aggregators, Speko is provider-neutral and benchmark-driven. It competes not by offering its own models, but by providing empirical data and intelligent routing across the best models from others (e.g., AssemblyAI, OpenAI, Google). This contrasts with vendors who only benchmark their own models against limited public datasets.
2. Key Innovation: The dual-architecture of a managed Router plus an open Gateway. This provides a seamless path from prototype (using the managed service) to scaled production (using the Gateway for direct, controlled connections). The public commitment to continuous, multi-language benchmarking creates a transparent, data-centric foundation for routing decisions that is unique in the market.

Frequently Asked Questions (FAQ)

1. How does Speko's routing actually work? Speko's Router uses a decision engine informed by its live benchmark database. When an API request is made (specifying language, modality, and optional objective like "highest_accuracy"), the system evaluates the current performance, latency, and cost rankings of all integrated providers for that specific context and routes the request to the optimal endpoint.
2. What is the difference between Speko's Router and Gateway? The Router is a fully hosted API where Speko manages all provider connections, keys, and routing logic. The Gateway is an open-source component you deploy; it allows your application to connect directly to providers using your keys, while still optionally using Speko's intelligence for routing, offering greater control and reduced latency.
3. Does using Speko create additional latency in my voice AI pipeline? For the hosted Router, there is minimal added network latency as it operates as an optimized proxy. For the Gateway architecture in a real-time streaming scenario, latency can be lower than a manual integration, as it establishes efficient, direct WebSocket connections to providers and handles routing logic locally.
4. How does Speko make money if it doesn't charge for the Gateway? Speko operates on a usage-based SaaS model for its managed Router service. Customers are billed for the AI inference consumed through the Router's unified API. The open-source Gateway is free to use, incentivizing adoption and allowing customers to choose their preferred architecture.
5. Are Speko's benchmarks reliable and unbiased? Speko's benchmarks are presented as an independent, continuous measurement system. Their methodology is public-facing, showing scores across many languages. As a router, their incentive is to identify the truly best-performing model for each scenario to ensure customer success, aligning their goals with accurate data.