The listening layer for voice agents
Oruk is the listening layer for voice agents. Its Realtime API streams transcription in 32 locales over one WebSocket and scores the emotion in each phrase, so the agent hears how a caller sounds as well as what they said. You bring the language model, the speech synthesis and the telephony.
Updated
What Oruk does in the loop
One of four parts. The other three stay yours.
| Part | Job | Provided by |
|---|---|---|
| Listening | Transcribe the caller and score the tone of each phrase | Oruk Realtime |
| Reasoning | Decide what to say and which tool to call | Your language model |
| Speaking | Turn the reply into audio | Your speech synthesis |
| Transport | Carry audio to and from the caller | Your WebRTC or SIP provider |
What arrives, and when
| Event | Carries | Use it to |
|---|---|---|
conversation.item.input_audio_transcription.delta | Provisional words | Show live text; detect that the caller is still talking |
conversation.item.input_audio_transcription.completed | Final transcript for the phrase | Send the turn to your language model |
conversation.item.input_audio_emotion.completed | Phrase timestamps, text, phrase ID and emotion scores | Add tone to the prompt, or change pacing |
conversation.item.input_audio_emotion.failed | A phrase that could not be scored | Carry on without tone; do not assume neutral |
A phrase ends after 600 ms of silence by default, adjustable from 200 to 2,000 ms. Shorter boundaries respond sooner and cut more phrases short. Measure on your own microphones and network before setting a latency target.
Give the model the tone
The words alone lose the delivery. Pass both.
{
"role": "user",
"content": "Fine. Just cancel it then.",
"voice": { "frustrated": 0.81, "angry": 0.42, "calm": 0.06 }
}Illustrative values. Scores are independent, do not sum to one, and describe the sound of the phrase.
What it costs
Realtime audio draws on a plan’s speech understanding minutes: one minute of audio uses one minute, with or without live speaker labels. Plans start at $9 a month with a 7-day trial.
If you only need final text for short English utterances, hosted Orukeet transcribes at $0.00045 a minute. Estimate a monthly bill.
Questions
- Does Oruk run the whole voice agent?
- No. Oruk transcribes the caller and scores their tone. The language model, speech synthesis, turn-taking policy and telephony are yours, usually through a framework such as Pipecat or LiveKit Agents.
- What does the agent receive?
- Provisional words as they are recognised, a final transcript for each phrase, and a separate event with that phrase’s emotion scores. The emotion event arrives after the words, so join the two by phrase ID.
- How long can a session last?
- Ten minutes per Realtime session. Open a new session for a longer call.
- Which audio formats work on a phone call?
- Realtime takes mono signed 16-bit PCM at 16 kHz. Telephone audio arrives as 8 kHz mu-law, so decode and resample it on your server before sending. The Twilio guide shows the conversion.
- Do the scores say how the caller feels?
- No. They describe how the speech sounds. Use them to adjust pacing or to flag a call for review, never as the only basis for a decision about a person.
Try it on your own audio.
The demo needs no account. The trial needs a card.