Launch HN：Speko（YC S26）——面向语音 AI 的 OpenRouter。

Backed by Y CombinatorBacked by Y CombinatorBacked by Y Combinator→The Router for Voice AI
Every speech model, benchmarked language by language, wired into one API.
Get API keyCompare models

Router
Which model to call, per language and per objective, decided from published measurements instead of a vendor's English leaderboard.

Benchmark coverage by language
Speech-to-textSpeech-to-text
Text-to-speechText-to-speech

ModelENARFRDEHINBESTATE
Nova-3

stt-rt-v5

GPT-4o Transcribe

GPT-4o-mini Transcribe

Qwen3-ASR

Chirp 3

Ink-Whisper

Pulse Pro

Whisper-1

Nova-2

Gradium ASR

Grok STT

Ink-2

Pulse

Realtime STT-1bench

Scribe v2 Realtime

Solaria-1

Universal-3.5 Pro

Velma 2

Flux

GPT Live Transcribe

Show all 21 models↓

Score against cost, per stage
Speech-to-textSpeech-to-text
LLMLLM
Text-to-speechText-to-speech
Speech-to-speechSpeech-to-speech

Accuracy (WER)
$0.0010$0.01602.0%11.0%Realtime STT-13.3% · $0.0025
Universal-3.5 Pro2.0% · $0.0075
Velma 24.4% · $0.0010
GPT-4o Transcribe2.3% · $0.0060
GPT-4o-mini Transcribe2.7% · $0.0030

Cost ($/min)

#ModelWER▲batchCost▼/ min
1Universal-3.5 Proassemblyai:universal-3-5-pro
2.0%$0.0075
2GPT-4o Transcribeopenai:gpt-4o-transcribe
2.3%$0.0060
3GPT-4o-mini Transcribeopenai:gpt-4o-mini-transcribe
2.7%$0.0030
4Qwen3-ASRalibaba:qwen3-asr-flash
2.8%$0.0054
5Realtime STT-1inworld:inworld-stt-1
3.3%$0.0025
6Chirp 3google:chirp_3
3.9%$0.0160
7Velma 2modulate:velma-2-stt-streaming-english-v2
4.4%$0.0010
8Grok STTxai:stt
4.8%$0.0033
9Solaria-1gladia:solaria-1
5.0%$0.0125
10Pulsesmallest:pulse
5.1%~$0.0050
11stt-rt-v5soniox:stt-rt-v5
7.5%$0.0020
12Gradium ASRgradium:default
8.4%$0.0104
13Nova-3deepgram:nova-3
9.8%$0.0048
14Ink-2cartesia:ink-2
11.0%$0.0090
—Fluxdeepgram:flux-general-en
—$0.0065
—Scribe v2 Realtimeelevenlabs:scribe_v2_realtime
—$0.0065
—GPT Live Transcribeopenai:gpt-live-transcribe
—$0.0170

Show all 17 models↓

Gateway
One base URL and one key in front of every provider, speaking the API your framework already calls.

Run and observe your voice workers
Speko speaks the OpenAI API, so the frameworks you already use need a hostname and a model string.

- LiveKitTypeScriptPoint the OpenAI plugin at the gateway. The agent code is unchanged.

- PipecatPython

- Claude / CursorMCP
copyimport { defineAgent, voice } from '@livekit/agents';
import * as openai from '@livekit/agents-plugin-openai';

const key = process.env.SPEKO_API_KEY!;
const baseURL = 'https://api.speko.ai/v1';

export default defineAgent({
entry: async (ctx) => {
const session = new voice.AgentSession({
stt: new openai.STT({ apiKey: key, baseURL, model: 'auto' }),
llm: new openai.LLM({ apiKey: key, baseURL, model: 'auto' }),
tts: new openai.TTS({ apiKey: key, baseURL, model: 'auto' }),
});
await session.start({ agent: new voice.Agent({ ... }) });
},
});

Point your agent at Speko

MCPMCPcurlcurl

$claude mcp add --transport http speko https://mcp.speko.ai/mcp

Read the docsGet API key