Skip to content

Belhulp

Listens in on sales calls, transcribes live and fills in the call form by itself.

Call assistant for the MacIn use during calls

fig. VIprovisional · Whisper small, instantfinal · Gemini, per windownowwindow runningreplaces0 s45 s90 s135 sone window
fig. VI Grey provisional text, replaced by the final version every 45 second window.

Who it is for

Myself, for the sales calls I make for school.

The problem

During a sales call you have to listen, ask follow-up questions and keep a form up to date, all at once. Typing it up afterwards takes time, and names, phone numbers and agreements get lost or end up wrong.

What I built

Belhulp runs on my Mac. My voice and the customer's are recorded separately and transcribed straight away. From the conversation it fills in the questions of the call form, keeps every call in an overview, and uses a browser extension to put the answers into the system I work in.

What it delivers

  1. No more typing up after a call.

  2. Names, numbers and agreements end up correct.

  3. Every call is neatly kept in one overview.

  4. Without the customer's consent nothing is recorded or stored.

Under the hood

  • Fast speech recognition on the Mac itself for the live view, and a more accurate model that corrects every part of the call afterwards.

  • Fields are filled through a fixed schema with an 'unknown' option, so nothing is guessed.

  • Every call goes to a queue first, so none is ever lost.

The hardest part

Bits of audio dropped out whenever speech recognition claimed the processor, which mangled names. I separated recording from processing, so the recording never has to wait. After that no audio was lost.

Built with

  • Python
  • faster-whisper
  • Gemini
  • Flask
  • Swift
  • Chrome-extensie