Voice
The agent can say it out loud.
Acumane Voice is speech for the agents you already run. It listens, answers in a chosen voice, and can place a call. The engine is Resemble AI's Chatterbox-Turbo: a compact English model with a one-step decoder, voice cloning from a short clip, and tags such as [laugh] and [chuckle].
Real Chatterbox audio keeps Resemble's Perth watermark. Acumane does not strip it. When no GPU service is connected, the same screens run on a local tone so a practice can try the desk, the consent gate, and the call log before any vendor account exists. That tone is marked as development audio. It is not Perth, and it is not a person's voice.
01
A desk you can hear
Push to talk, or leave the microphone open. The transcript arrives as you speak, and the reply starts as audio. Talk over it and it stops.
02
Voices that stay in the firm
Harbor, Fieldnote, and Northwind are Acumane timbres. They are not recordings of a person. A clone is built from a reference clip and cannot be selected until the owner records the consent script for this organization.
03
A call that introduces itself
The first sentence is a disclosure: this is an AI assistant, and the call is recorded. You confirm the number. An admin approves the call. The number belongs to your own Twilio account. Twilio sends the invoice for the minutes.
Metered like the rest of the work
Spoken characters and transcribed minutes debit the credit ledger. The proposal is three cents per thousand characters and six tenths of a cent per transcribed minute, with the same twenty percent margin as managed model calls. Telephone minutes are not in that ledger. You connect the Twilio account your firm already pays, and Twilio bills those minutes directly. A key you bring for Resemble, ElevenLabs, OpenAI speech, or Deepgram records the provider cost and does not spend credits.
Voiceovers download as WAV, and as MP3 when the host has ffmpeg. With object storage configured, a copy is kept in the bucket. A finished call writes a transcript, a summary, a recording, and a task result into ARSI.
What a real deployment still needs
Speech from the model needs the GPU image and the Chatterbox weights. A live phone call needs the Twilio account you connect in the workspace, a number on that account, and a public WebSocket address. Acumane does not lend a shared Twilio account. Until a customer account is connected, an approved call is simulated and still recorded.
Chatterbox-Turbo is MIT licensed. The deployment notes are in the voice guide.