Congrats to @GradiumAI on open-sourcing Phonon. Phonon-31m is out under Apache 2.0, small enough to run on device.
A new frontier for on-device speech.
We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date.
Phonon-31m is available under Apache 2.0.
gradium.ai/phonon🧵
Grok Voice 3.0 by @xai is live on Speko.
It answers the caller in 1.6s, the fastest of any S2S model we time, and has the top task capability on our board: 0.94.
Try it: platform.speko.ai
Speko Translate is now live!
translate.speko.ai
Introducing Speko Translate.
Speko Translate calls any phone number from your browser and interprets the call live in both directions, in 51 languages, using the Speko calling API and the Speko Router.
Never put off a phone call because of the language again.
Try it now at translate.speko.ai
The @speko_ai MCP server is #1 in Developer tools on the Manufact MCP catalog.
99/100, grade A+.
Connect it to Claude, ChatGPT, or Cursor and build voice agents from your editor: speko.ai/developers
The fastest Thai voice in Asia.
Paxa Labs starts speaking in 89ms from Singapore. The next fastest takes 177ms. ElevenLabs takes 789ms.
It's 2nd of 14 on getting the tones right, too.
On Speko now. benchmarks.speko.ai
You can now bring your Twilio numbers to Speko.
Paste your Account SID and API key. Speko builds the SIP trunk in your Twilio account and attaches your numbers. No forwarding, no second console.
Try it: platform.speko.ai
Palabra TTS is live on Speko, at 50% off.
Realtime speech synthesis built for live translation and dubbing, now routed through one API alongside 15 other text-to-speech models.
Try it: speko.ai/models
ElevenLabs v4 and v4 Turbo just posted the highest robustness we have ever measured on our TTS board
v4 also leads on naturalness.
Turbo trades some naturalness for 205ms latency at half the price.
See the board: benchmarks.speko.ai/tts
Both are now routable on Speko. Pick eleven_v4 or eleven_v4_turbo on platform.speko.ai, same as every other model on the gateway.
Stop guessing which speech model is fastest or cheapest.
Pick up to 3, set your language, and see latency, price per minute, and capabilities next to each other.
Speech to text, text to speech, and language models.
Try it: platform.speko.ai
Gemini 3.8 Flash TTS by @GoogleDeepMind is the most robust voice on our text-to-speech board.
We sent "-$12.50". It said "negative twelve dollars and fifty cents".
0.95 on reading numbers, dates and money right.
Full Result: benchmarks.speko.ai/blog/gem…
We measured Gradium TTS Beta on our TTS board: 49ms p50, 56ms p90 time to first audio from US East.
That is more than 4x faster than the previous Gradium TTS (217ms p50), with the same naturalness.
Full results: benchmarks.speko.ai/tts
Gradium TTS beta model is out now to define what fast means.
Less than 50ms TTFA, same quality. Most models are >100ms TTFA.
Try it now: studio.gradium.ai/?model_nam…
Jev by @typesafe_ai now decides which branch a Speko workflow takes.
It scores every option on the node, so it can't invent a branch. A language model can.
It got nearly all our test routes right, and runs 40-60% faster than our fastest LLM.
Try it: platform.speko.ai
Speko CLI is Live!
Your coding agent can now build and test voice agents without touching a dashboard:
npm i -g @spekoai/cli
speko-cli login
Jev by @typesafeai now routes through Speko.
It returns typed decisions with calibrated confidence on every turn, so your agent can branch mid-call on intent, handoff, or whatever you define.
One call per turn. No parsing.
Try it: platform.speko.ai
Gemini 3.8 Live by @GoogleDeepMind is live on Speko.
It takes the top row on our S2S board: 0.92 task capability and a 2 second answer across 18 concierge calls.
Try it: platform.speko.ai
See details: benchmarks.speko.ai/blog/gem…
OpenAI shipped GPT-Live-1 this week.
We had been benchmarking it since the alpha, with their team.
It splits a voice agent in two: one model holds the conversation, another does the thinking.
It lands first of ten rows on our S2S board.
benchmarks.speko.ai/blog/gpt…