speko
start free trial

voice ai · routing intelligence

every provider.
one winning route.

one api across every leading voice provider. benchmarked continuously, routed per language and intent. failover under 500ms, mid-conversation. for teams that can't ship a misheard word.

start free trial→see live benchmarks
backed by y combinator · s26
speko-router · prod-us-west-2
last update 0s

routed

elevenlabs·scribe-v2 → cartesia·sonic-3

WER

3.5%

latency

145ms

confidence

97%

live scoring · 6 providers

  • 1elevenlabs·scribe-v298
  • 2alibaba·qwen3-asr-flash98
  • 3assemblyai·universal-3-pro96
  • 4google·chirp-295
  • 5elevenlabs·scribe-v195
  • 6google·gemini-2.5-flash94

failover queue

  • ⤷elevenlabs·scribe-v2 → alibaba·qwen3-asr-flashp99 spike · 412ms
  • ⤷elevenlabs·scribe-v2 → assemblyai·universal-3-prorate limit · 287ms
live
routes/sec1,284
p99 latency412ms
failovers/min3.2
uptime99.998%
updated 0s ago

Voice AI · Routing intelligence

Stop guessing which voice stack to use.We tested all of them.

Live English benchmarks across every major STT, LLM, and TTS provider. Pick an agent below or describe your own — get the ranked stack and test it in your browser before you ship.

build an agent for

or describe your own

Your recommendation will appear here.

Enter a use case above and hit Optimize.

how it works

benchmark. route. failover.

step 01→

benchmark

every provider, on the same audio. WER · latency · cost · noise robustness · diarization. updated daily, never with vendor demos.

26 datasets · 10+ languages
step 02→

route

each request scored against your intent — language, latency budget, cost ceiling. the winning provider gets the call, every time.

p99 routing decision · 4ms
step 03

failover

rate limits, regional outages, model regressions — speko detects and reroutes mid-conversation. your user never hears a hiccup.

failover · under 500ms

Plug in

Drop Speko into your stack in one line.

CLI for benchmarking from your terminal. MCP server for asking Claude to pick a stack and wire it up.

$

live demos

four agents. routed live. your call.

each one talks through a different speko-routed stack — same api, different winning provider. tap call, talk, hear the difference.

ready

sasha

vela cloud · support

language en · optimizeFor latency
via deepgram·nova-3 → cartesia·sonic-3
ready

miles

harbor health · outbound

language en · optimizeFor balanced
via assemblyai·universal-2 → elevenlabs·eleven_flash_v2_5
ready

amelia

the lemon hotel · reception

language en · optimizeFor quality
via elevenlabs·scribe_v2 → elevenlabs·eleven_v3
ready

halo

northwind utility · ivr

language en · optimizeFor cost
via openai·gpt-4o-mini-transcribe → cartesia·sonic-3

faq

questions we get on twitter.

  • openrouter and portkey route LLM tokens. speko routes voice — STT, LLM, and TTS together — and it picks based on its own published benchmarks, not on what the provider tells you. we're not another voice provider. we're how you pick one.

ship accuracy you can audit.

one api. every provider. routed continuously, failed over mid-call, benchmarked in the open. for teams whose voice agent can't mishear a word.

start free trial→see benchmarks