Dialt retweeted
🔥 New on VAmoS Bench 🔥
Dialt Fast and Dialt Smart from @dialt_ai
41.1% pass rate and $0.18 per call, Dialt Fast is on the pareto frontier 🏅
it just bumps off both @OpenAI GPT-Live 1 and @GoogleDeepMind Gemini 3.8 Live
details -> veris.ai/leaderboard/vamos-p…
🧵/
Dialt retweeted
This voice agent was EXTREMELY RUDE!!!
..available now through the Dialt.com api, for approved orgs only.
Why is this useful?
- for training (e.g. simulated customers/patients)
- testing out your production agents
For custom simulations/characters reach out to hello at dialt dot com
Dialt retweeted
Today we’re launching Dialt.
Tune how your voice assistants sound, speak and act, down to the model weights.
We serve our own conversational voice models and train them on your data:
• Voice and accent
• Conversational style
• When and how they call your tools
Build a simulated patient who’s reluctant to share, a customer who pushes back, or an assistant that speaks your customers’ dialect.
Try the playground: dialt.com
And follow us @dialt_ai
Dialt retweeted
We are launching Dialt!
Join us on Wednesday for our official launch!
5 pm Dublin, 12 pm New York, 9 am SF.
Sign-up to attend: luma.com/3lygllj4
Now available through Pipecat!
Dialt's conversational voice engine is now also available through @pipecat_ai!
If you're using Pipecat, try flipping over to Dialt Fast or Dialt Smart and trying them out.
58
Dialt Fast and Dialt Smart - now with even more consistent fast responses!
Dialt Fast and Dialt Smart models have extended their lead over Gemini 3.8 Live and OpenAI’s GPT Live 1 models, with further improved response times (latency).
This is an internal cross-customer benchmark, but, as we like to say at Dialt…
“The most important benchmark the one you build yourself, for your own application.”
The Dialt API allows you to:
1. Quickly spin-up a benchmark that is bespoke to your use case
2. Measure performance of Dialt vs Gemini vs OpenAI models using that benchmark
3. [Coming soon] Create training data and tune Dialt models to further improve performance. Reach out for more info
1
51
These tips apply whatever your approach to building voice agents, @ElevenLabs , @pipecat_ai , @livekit etc.
At Dialt, they are built in as defaults.
More to come soon on how we think about evals.
3 tips for voice agent evals:
1. Test in text mode first. Cheaper, and faster than real time.
2. Use a strong simulated caller. If the caller is weak, you won't know if your assistant or the simulated caller was the problem.
3. Use a strong judge, with clear success criteria and a rubric.
1
74












