Speko launches a router for voice AI models
Speko, backed by Y Combinator's S26 batch, launched on Hacker News describing itself as "the router for voice AI." The pitch is that every speech model gets benchmarked language by language, then wired into one API, so the choice of which model to call for a given language and objective is decided from published measurements rather than a vendor's own English-language leaderboard.
The launch page publishes its own benchmark comparing about 20 speech-to-text (STT) models on word error rate (WER) against cost per minute. AssemblyAI's Universal-3.5 Pro tops the accuracy ranking at 2.0% WER for $0.0075 per minute. OpenAI's GPT-4o Transcribe follows at 2.3% WER for $0.0060 per minute, and GPT-4o-mini Transcribe at 2.7% WER for $0.0030 per minute. Alibaba's Qwen3-ASR scores 2.8% WER at $0.0054 per minute, and Inworld's Realtime STT-1 scores 3.3% WER at $0.0025 per minute. Google's Chirp 3 posts 3.9% WER at $0.0160 per minute, the highest listed cost among models with a WER figure. Modulate's Velma 2 is the cheapest model in the table at $0.0010 per minute, with 4.4% WER. Further down the list: xAI's Grok STT at 4.8% WER for $0.0033 per minute, Gladia's Solaria-1 at 5.0% WER for $0.0125 per minute, Smallest's Pulse at 5.1% WER for roughly $0.0050 per minute, Soniox's stt-rt-v5 at 7.5% WER for $0.0020 per minute, Gradium ASR at 8.4% WER for $0.0104 per minute, Deepgram's Nova-3 at 9.8% WER for $0.0048 per minute, and Cartesia's Ink-2 at 11.0% WER for $0.0090 per minute. Three more models appear with a cost but no WER score: Deepgram's Flux at $0.0065 per minute, ElevenLabs' Scribe v2 Realtime at $0.0065 per minute, and OpenAI's GPT Live Transcribe at $0.0170 per minute. The benchmark table's column headers also list nine languages (English, Arabic, French, German, Hindi, Norwegian Bokmal, Spanish, Tamil, Telugu), though the individual per-language, per-model scores are not present as readable data in the page's scraped text, only the headers and model names.
On the integration side, Speko speaks the OpenAI API, so an app already built on OpenAI-compatible STT, LLM or TTS clients only needs a new hostname and model string to route through it. A sample snippet shows LiveKit's voice agents SDK pointed at https://api.speko.ai/v1 with a Speko API key, using model: 'auto' for speech-to-text, the LLM and text-to-speech, letting the router pick the underlying model per call. The launch page also lists an MCP server, addable with the command "claude mcp add --transport http speko https://mcp.speko.ai/mcp", so agent frameworks with MCP support can reach the router directly.
The launch text gives no founder names, team size, or company location, and states no funding figure beyond being Y Combinator backed. It also does not give a date for either the YC batch admission or the product launch itself; the "S26" batch label appears only in the story's title, not in the page text.
Key facts
- Speko, backed by Y Combinator's S26 batch, launched on Hacker News as a router that picks which speech-to-text model to call per language and per objective, based on its own published benchmarks rather than a vendor's English-only leaderboard.
- Its benchmark scores roughly 20 STT models on word error rate against cost per minute; AssemblyAI's Universal-3.5 Pro leads at 2.0% WER for $0.0075/min, with OpenAI's GPT-4o Transcribe close behind at 2.3% WER for $0.0060/min.
- Modulate's Velma 2 is the cheapest model listed at $0.0010/min, though its WER is 4.4%; three models (Deepgram Flux, ElevenLabs Scribe v2 Realtime, OpenAI GPT Live Transcribe) are shown with a price but no WER score.
- The service exposes a single OpenAI-API-compatible endpoint (https://api.speko.ai/v1) so existing STT/LLM/TTS clients, including LiveKit's voice agents SDK, can route through it by changing only the hostname and model string, plus an MCP server for agent frameworks.
- The launch text names no founders, team size, location or funding amount beyond being Y Combinator backed, and gives no date for the launch or the YC batch admission.
Why it matters
Voice AI stacks today mean picking one speech-to-text vendor and living with its blind spots across languages and price points, usually judged against benchmarks the vendor itself publishes, in English. Speko applies the OpenRouter idea, a single API in front of many models with routing decided by measurement, to speech models specifically, adding a per-language angle that a single-vendor leaderboard cannot offer.
Who it affects
Developers building voice agents, real-time transcription or voice assistants who currently hardcode one STT provider and would otherwise have to benchmark and integrate each alternative themselves. Teams already using OpenAI-compatible tooling are the easiest fit: the sample integration is LiveKit's voice agents SDK, and switching to Speko is shown as a hostname and model-string change rather than a rewrite.
How to use it
Point existing OpenAI-compatible STT, LLM or TTS clients at https://api.speko.ai/v1 with a Speko API key; setting model: 'auto' lets the router choose the underlying model per call. An MCP server is also available, added with "claude mcp add --transport http speko https://mcp.speko.ai/mcp" for agent frameworks that speak MCP. No subscription price or plan tiers are given beyond the per-model, per-minute costs shown on the benchmark page itself.
How solid is it
The evidence for the routing claim is Speko's own benchmark table, not an independent evaluation. It is more specific than a typical launch announcement, naming about 20 models by exact provider identifier with WER and per-minute cost figures rather than vague marketing language, which gives it more to check than most launch posts. The underlying test methodology, dataset, and the actual per-language scores behind the nine listed languages are not visible in the page text used here.
Risks and caveats
A router that benchmarks and then routes traffic to the vendors it benchmarks carries the same neutrality question voice-AI teams already have about vendor-published leaderboards. The launch text gives no founder names, team size or company location, and states no funding amount beyond being Y Combinator backed; it also does not give a date for the YC batch admission or the launch itself, so how new or how funded the company is beyond that one line is not established here.
“Which model to call, per language and per objective, decided from published measurements instead of a vendor's English leaderboard.”
— Speko launch page