About Services Tools Blog Contact us

Always-On Speech Transcription

A background listener that records speech from the main microphone and saves each phrase as timestamped English text, with a health check and start and stop buttons.

  • Speech Recognition
  • Professional services
  • Scheduled worker
  • Product
  • Local/private
  • 2026

The problem

Spoken notes and asides vanish unless someone types them up. Dropping audio files in by hand does not suit a microphone that should keep listening all day.

What it does

A background program started in the signed-in session detects when someone is speaking and writes each phrase out as text with Whisper. It stores the time, text, language and a confidence score. A watcher restarts it if its health check goes quiet, and the screen can start or stop it without the website holding the microphone.

Outcomes

  • A running log of what was said, with no manual uploads.
  • Low-confidence lines can be reviewed instead of trusted blindly.
  • A stopped listener is restarted by the health check.
  • Start and stop work from the screen, and the website never holds the microphone.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Desk listener Listening

Utterances today
0
Heartbeat
0
fresh
Needs review
0
Language
EN
local model
Default microphone
Energy detector on Two voices, one clip

Play reads the sample. This page never opens the microphone.

Sample replay
Ready to play the desk clip
Utterances
SpokenTextSecClipLang
0:00 Book the boardroom for Thursday at 3. 2.5 en
0:03 The mezzanine rack is short two cartons of the Navy court. 4.0 en
0:08 Send the packing note to the studio buyer before noon. 3.0 en
0:13 Move the north and delivery to Friday morning. 3.2 en
0:17 Hold the courier until the labels are printed. 2.6 en
0:21 I will ring the dispatch desk after the morning meeting. 3.8 en

Low-score review

Under 0.60

These lines are kept, not dropped. A person accepts or retypes them. The 0 to 1 figures are a review stand-in on this page. The live store keeps an average log probability.

Waiting
0
Accepted today
0
Lowest score
0.42
Queue
SpokenTextScoreState
09:06:03 Something about the hem on the sample. 0.42 needs review
09:18:44 Put it on the second rail, not the window. 0.51 needs review
08:41:12 Invoice for the quay workshop, net thirty. 0.58 accepted
09:27:05 Check the burgundy swatch against the window light. 0.47 needs review

Heartbeat supervisor

4s old The listener is writing a heartbeat on every captured frame. Simulate a stall to see the restart.
Watchdog log
Idle, last check 30s ago
Beats this morning

The dip at 12 is the stall this preset replays.

How it works

  1. Listen to the main microphone in the signed-in session.
  2. Cut out each spoken phrase by detecting when someone is speaking.
  3. Turn each phrase into English text with Whisper.
  4. Save the time, text, language and confidence score.
  5. Restart if the health check goes quiet, and follow start, stop and restart from the screen.

Pipeline

flowchart LR
  Mic[Default microphone] --> Vad[Voice detector]
  Vad --> Buf[Utterance buffer]
  Buf --> Asr[Whisper]
  Asr --> Row[Timestamped row]
  Row --> Beat[Heartbeat]
  Ui[Start or stop] --> Mic

Sample output

Short two-voice desk exchange, with the local transcript.