Every leading voice model, on a real phone line
Put OpenAI Realtime, Gemini Live, ElevenLabs, or Deepgram on your phone numbers. DialPress carries the call, the audio, and the tools; you choose the model and switch it with one line.
Speech-to-speech and speech providers
At a glance
One line
Switch an agent’s model without touching its prompt, tools, or routing.
Same behavior
Tools, barge-in, handoffs, and events work the same on every model.
Your keys
Use your own provider keys, per agent or for the whole account.
Or none
Run open models on your own GPUs when audio must stay in your network.
Speech-to-speech agents
One model hears the caller, thinks, and answers in its own voice.
01
OpenAI Realtime
gpt-realtime and the other realtime models, through the LLM gateway. A G.711 call’s audio goes to the model as it is.
02
Gemini Live
Gemini’s native audio models, from AI Studio or Vertex, connected directly with your Google credentials.
03
ElevenLabs Agents
ElevenLabs’ conversational agents on your numbers, with their voices.
04
Deepgram Voice Agent
Deepgram’s voice agent, with the thinking model of your choice.
05
Also
xAI, Azure OpenAI, Azure Voice Live, and Amazon Nova Sonic.
Or build the pipeline yourself
A cascade agent picks each part: speech recognition, a language model, and a voice. Mix Deepgram’s recognition with any model and an ElevenLabs voice. The caller’s turns are found by voice activity detection, and answers are spoken a sentence at a time, so the first words play while the rest is generated.
Speech providers
Pick any voice. Pick any ear.
Text to speech and speech to text from 28 providers, each a setting on the agent. Mix one company’s recognition with another’s voice.
Text to speech
Speech to text
Both, and agent speech
Speech recognition and voices
Every provider’s models and voices are listed live from its API, so new ones appear without an update.
Deepgram
Nova live recognition and batch transcription, with your vocabulary as key terms. Aura 2 voices.
ElevenLabs
Scribe live and batch transcription. Flash, Multilingual, and v3 voices, including your account’s cloned voices.
OpenAI
Live transcription with gpt-4o-transcribe. gpt-4o-mini-tts voices with speaking instructions.
Google Cloud
Chirp 3 recognition, with streams that run past five minutes. Chirp 3 HD and Neural2 voices.
Azure AI Speech
Recognition and fast transcription. Neural voices with SSML style and prosody.
Amazon
Transcribe streaming. Polly neural, generative, and long-form voices.
Carriers and messaging channels
Carriers and messaging channels
Reach callers and customers where they already are.
Twilio
Connect Twilio Elastic SIP Trunking as a trunk. Your Twilio numbers ring your agents and apps, and calls out leave through Twilio.
WhatsApp conversations reach the same chat agents, knowledge base, and handoffs as your calls and SMS.
Telegram
Telegram chats answered by the same agents, with a person taking over when the agent should stop.
Language models
Every model call goes through the LLM gateway.
Any provider
Pick from your account’s model catalog: OpenAI, Gemini, Anthropic, and the rest.
Budgets per call
Each call gets its own key with a budget, so one runaway conversation can’t run up the bill.
Costs you can see
Cost per day by agent and by model, and the usage of any single call.
Prompts with versions
Version, compare, and promote prompts, with traces of every turn.
Open models on your GPUs
For audio that must not leave your network.
Recognition
Streaming open models such as Whisper and Zipformer, balanced across your GPU machines.
Voices
Kokoro and Piper voices, generated next to the call.
Compatible servers
Point the OpenAI integration at a self-hosted compatible server and keep the same setup.
Common questions
Can I switch models without rebuilding my agent?
Yes. The model is one setting. Prompt, tools, routing, and handoffs stay the same.
Do I need my own provider keys?
You can use your own keys per agent or per account. Speech is also included in DialPress cloud.
Do tools and transfers work on every model?
Yes. Built-in phone tools, your MCP tools, barge-in, and handoffs to people behave the same on speech-to-speech and cascade agents.
Can I use a model you don’t list?
Any model in your LLM gateway’s catalog works as the language model, and OpenAI-compatible speech servers work through the OpenAI integration.
Built from
The DialPress primitives this uses.
Put your model on a phone number.
Pick a model, give it your tools, and call it.