Speko Launches as an OpenRouter for Voice AI. What Changes for Voice Stacks

GPTLive1 is an independent site. We are not affiliated with, endorsed by, or connected to OpenAI or Speko. This is an explainer, not an official source.

A Launch HN post introduces Speko, a Y Combinator Summer 2026 company that calls itself an OpenRouter for voice. The public pitch is one API key in front of speech-to-text, an LLM, and text-to-speech, with routing driven by language-specific benchmarks rather than a vendor's English-only leaderboard.

That is a real layer in the voice stack. It is not a full-duplex conversation model, and it is not an audiobook studio. For readers of GPTLive1, the useful split is: Speko chooses which speech and synthesis endpoints to call. GPT-Live-1, as OpenAI describes it, is a model class for full-duplex voice. A free in-browser TTS tool is a third thing: a way to hear a script without standing up an agent.

This article stays inside what Speko and YC have published, plus the distinctions that keep those layers from being mashed together in a blog post.

What Speko says it measures

Speko's launch materials and docs on scoring describe a continuously running benchmark suite. Rankings are per language and region, and they differ by modality:

• STT: word error rate, time-to-first-partial or finalize latency, cost per minute.
• TTS: a character-error or naturalness-related score, time-to-first-byte, cost per minute (with a conversion for character-billed providers).
• LLM: a quality score, time-to-first-token, blended token price.
• Speech-to-speech: task success and tool-call latency where that path exists.

They publish boards at benchmarks.speko.ai and say a new vendor model is tested the same day. Routing can change when the snapshot changes; two identical calls inside one snapshot should pick the same stack. A scoresRunId on the decision is how you debug "why did we get provider X this morning?"

That design copies OpenRouter's aggregator idea — one key, many backends, fallback on failure — and applies it to the speech path, where English leaderboards routinely mis-rank Spanish medical dictation or a ten-minute unscripted take.

The launch HN comment thread adds operational color that should be treated as reported, not independently audited here: an MIT-licensed gateway for teams that do not want an extra hop or do not want to share keys; BYOK mode that Speko says does not phone home; anonymous telemetry on by default with an env var to disable it. Read the GitHub repo and the privacy toggle before you assume any of those claims for a production deployment.

What a voice router does not replace

It does not replace model selection for conversation. Full-duplex voice — overlapping speech, interruption, barge-in — is a different product surface from "transcribe this clip, then speak the answer." OpenAI's GPT-Live-1 line is the thing GPTLive1 exists to explain. A router can sit in front of STT and TTS while a duplex model handles the conversation. It cannot turn a half-duplex cascade into Live-1 by swapping vendors.

It does not replace script craft. TTS quality boards measure how a fixed utterance sounds. They do not fix a script that asks the engine to speak a URL, a nested parenthesis, or a heading that was meant for a page. If the input is a blog post pasted into a voice agent, the failure is editorial. See the site's TTS script formatting guide before blaming the provider.

It does not replace listening QA. Speko's own launch note distinguishes a 30-second "which clip sounds better" demo from production, which has to survive minute eight, money and dates, and spontaneous speech. Rankings change across those tests. A router that follows the board is only as good as the board's task. Your task may be a 40-minute narration, a call center, or a game NPC. Measure that.

It does not make vendor lock-in vanish. You still depend on whoever won the snapshot, on Speko's selector (or your copy of the open gateway), and on each provider's outage. Fallback is not the same as hedging, and neither is the same as a duplex model with a single websocket.

A practical way to use the launch without cargo-culting it

If you are building or buying a voice path this month:

1. Write down the job. Live agent, voicemail transcription, e-learning narration, and "read this email aloud" have different winners. Speko's language selector is the right instinct; a single "best TTS" is not.
2. Separate cascade from duplex. If you need interruption and overlap, start from a full-duplex model explainer, not from an STT→LLM→TTS router. GPTLive1's GPT-Live-1 overview is the duplex side of that split.
3. Pin a benchmark snapshot when you ship. If routing is allowed to move under you, record scoresRunId or the equivalent. Otherwise yesterday's QA is not today's production.
4. Keep a golden set. Ten clips in the languages you actually serve, including names, numbers, and a long take. Run them when the board updates. Public WER is not your WER.
5. Do not share keys by accident. If BYOK and a local sidecar matter, verify the gateway's network behavior yourself. A launch post is not a security review.

The in-browser TTS on GPTLive1 remains a way to hear a cleaned script without standing up LiveKit, Pipecat, or Speko. It is a listening aid, not a router.

How this launch fits the next twelve months

Voice stacks have been assembling from parts: a transcriber from vendor A, a chat model from vendor B, a speaker from vendor C, glue in Python. OpenRouter made that pattern normal for text models. Speko is attempting the same for speech, with the extra burden that quality is language- and task-shaped.

Expect three messy outcomes, none of which require believing every customer anecdote in the HN thread:

• More frequent provider swaps, because a board update is cheaper than a quarter-long R&D project. That is good if your tests catch regressions; bad if you have no tests.
• A new outage class: the router is healthy, the top-ranked STT is not, and fallback is a second-best model you never QA'd in that language.
• A documentation gap between "auto" as a model string and the actual model that spoke. Keep the snapshot id in logs.

None of that changes how full-duplex conversation works, and none of it changes how you format a voiceover script. It changes who you call for STT and TTS when you are assembling a cascade.

If you only remember one distinction: Speko is a routing and measurement company. GPT-Live-1 is a conversation model family. GPTLive1 is an independent explainer plus a small TTS demo. Mixing those three into one sentence is how a launch post becomes slop.