Turn detection
Deciding when the user is done. Too early interrupts, too late feels slow, so we tune it per language.
Hebrew & English AI voice agents
Voice agents that answer in time.
And stop when you interrupt.
Speech-to-text, a model with tools, and a voice, all streaming. Callers judge it in the first second, so we build and measure the whole chain on real audio in your languages.
Ultrabot, built by MLAIA Data ScienceLed by Dr. Yochai Edlitz · Ph.D., Weizmann Institute
The pipeline
Three models in a row.
Each one adds delay.
We stream every stage and start speaking at the first sentence, not after the whole reply.
Streaming speech-to-text
Chosen by accuracy on your vocabulary and streaming delay. Unclear fragments are dropped, not answered.
The agent
A model with tools for booking, lookup or updates, released one sentence at a time.
Text-to-speech
Voices compared on naturalness, pronunciation and time to first audio, with a fallback if one is slow.
Barge-in
The agent pauses when the user speaks. A cough resumes playback; a real request stops the turn.
Speech-to-speech models
Nova Sonic or realtime models listen and speak in one step. We compare them with a pipeline before choosing.
What we have built
A Hebrew and English voice assistant,
with an avatar that speaks.
For a healthcare app, currently in development: Amazon Transcribe for Hebrew in an EU region, Claude on Bedrock, sentence-by-sentence speech, soft barge-in, and a device-voice fallback when the cloud is slow. It builds on MLAIA’s audio and speech work, including noise reduction and echo cancellation.
We describe client projects in general terms. Details are shared in conversation, subject to confidentiality.
Hebrew & multilingual voice
Hebrew needs its own tests.
Not a translated English setup.
Missing vowels, gendered address, and callers who switch languages mid-sentence.
Pronunciation
Vowel marks and phonetic hints for names, tested on your vocabulary.
Recognition on your terms
Custom vocabularies, tested on recordings with your product names.
Gender & register
The right grammatical gender and register, in the text and the voice.
Voice stack optionsElevenLabs ScribeAmazon TranscribeDeepgramGoogle SpeechWhisperElevenLabsGoogle Chirp & WaveNetAmazon PollyCartesiaNova SonicWebSocket & WebRTCLiveKitAmazon ConnectTwilio
A practical delivery path
Measure on real audio.
Then tune the slowest stage.
Define the conversation
Callers, languages, channel and allowed actions.
Bake-off
Speech-to-text and voices compared on your recordings.
Build & tune
Streaming, turn detection and barge-in, tuned on test calls.
Operate
Delay, interruptions and handoffs monitored, with privacy rules.
Before we start
Good questions. Straight answers.
How fast can a voice agent respond?
It depends on the channel, models and network. We measure each stage on your setup and stream all of them, so the agent starts speaking at its first sentence instead of after the full reply.
Can we use ElevenLabs?
Yes, ElevenLabs is one of the voices we integrate and compare. For personal or health data, we check the provider's data terms first and use a provider with the right contract where needed.
Can a voice agent work over the phone?
Yes, through Amazon Connect, Twilio or a SIP provider. Phone audio is narrower-band and noisier than web audio, so recognition and turn detection are tested on real calls.
What happens when the agent cannot help?
It says so and hands off: transfer to a person, a callback request, or a message with a summary of the conversation so far, depending on the channel.
Do you record the calls?
Only if you decide to and the user is informed. Many of our designs keep no audio or transcripts at all and log only which tools were used.
Keep exploring
Related reading & tools.
Talk to a person
Tell us what
you’re working on.
Share the problem, the data you have and what success would look like. We’ll discuss a practical next step.
Please don’t include confidential datasets, credentials or patient information.