Speech recognition and text to speech, next to the call
Streaming recognition, text to speech, keyword spotting, and speaker recognition, in the same stack as the media path. Use open models on your GPUs or a cloud speech provider.
At a glance
Streaming
The caller hears the first words while the rest is still being generated.
Close
Speech runs beside the gateway, so audio doesn’t cross the internet to be transcribed.
Yours
Open models on your own GPUs, or a cloud speech provider.
Balanced
Requests go to the least-loaded healthy endpoint that serves the model.
Silence is the failure callers notice
A natural conversation leaves a few hundred milliseconds between turns. Recognition has to decide the caller is done, the model has to answer, and the first audio has to play. Keeping speech on the same infrastructure as the call removes network hops from every turn.
Capabilities
Speech recognition
Streaming transcripts with end-of-turn detection, for bots and live captions.
Text to speech
Sentence-by-sentence synthesis played straight into the call.
Keyword spotting
Hear a phrase the moment it is said, mid-sentence, and branch the call on it.
Speaker recognition
Tell the people on a call apart.
Run speech on your GPUs
Speech engines can run on GPU machines over mutual TLS. Weight bigger GPUs higher, and route each model only to the machines that serve it.
Bring the voice model you trust
Speech-to-speech or speech, model, and voice: switch with one line, keep the same tools and call handling.
Common questions
Can I use my own speech models?
Yes. Run open models on your GPUs, or use one of 28 speech providers, from Deepgram and ElevenLabs to Cartesia and AssemblyAI.
Can I use speech without an agent?
Yes. Recognition and keyword spotting work on any call, with or without an agent.
Where does call audio go?
To the speech endpoints you configure. On a self-hosted stack, it never leaves your network.
Transcribe your first call.
Speech is included in DialPress cloud, and runs on your GPUs when you self-host.