Who spoke when
Separate a recording by speaker so long meetings, interviews and calls are easier to follow and search.
Back to audio machine learning
AUDIO MACHINE LEARNING · POSSIBILITIES
What your sounds
are telling you.
Machines, phone calls, sites and spaces are full of useful sound that usually goes unrecorded. Audio machine learning turns it into alerts, transcripts and measurements you can act on. Here are some of the ways it can help.
01 / IN PRACTICE
Five ways to put
sound to work.
Each example below is a common starting point. Real projects are shaped around your own recordings, equipment and the decisions you want to support.
01 / MACHINE HEALTH MONITORING
Hear a fault before it becomes a breakdown.
A microphone or vibration sensor beside a motor, pump or fan learns what healthy running sounds like. A worn bearing, a slipping belt or a blocked filter changes that sound slightly, often days or weeks before anything fails, so repairs can be planned instead of rushed.
- Motors, pumps, compressors & fans
- Bearing wear, imbalance & misalignment
- Maintenance planned around real condition
Fewer surprise breakdowns, and repairs on your schedule.
Machine health monitoringILLUSTRATIVE DEMO
FREQUENCY · LAST 60 MIN
ANOMALY SCORE
CONDITION
Normal running Bearing wear likely 0.91
- 3.2 kHz tone rising
- Matches bearing wear pattern
- Inspect within 14 days
WORK ORDER RAISED BEFORE FAILURE ~14 days notice
02 / CALL TRANSCRIPTION & NOTES
Turn calls and meetings into notes and actions.
Speech recognition turns recorded calls, voicemails and meetings into searchable text, labelled by speaker. A language model can then summarise what was agreed and pull out the follow-up actions, so nothing relies on memory or a scribbled note.
- Customer calls & voicemail
- Site meetings & interviews
- Summaries, actions & CRM notes
Every conversation captured, without anyone taking notes.
Call transcription & notesILLUSTRATIVE DEMO
- CallerHi, the boiler at Unit 6 is banging again.
- YouIs that the one we serviced in March?
- CallerYes. Could someone come out Thursday?
- YouThursday at nine is fine. I’ll send a quote.
SUMMARY
ACTIONSRepeat fault on the Unit 6 boiler, last serviced in March.
Book engineer · Thu 09:00
Send quote to caller
NOTES SAVED TO CUSTOMER RECORD 2 speakers
03 / SOUND EVENT DETECTION
Recognise the sounds that need a response.
Models can be trained to pick out particular sounds from background noise — breaking glass, alarms, a machine stopping or a vehicle reversing — and raise an alert or log the moment it happened. It works where cameras can’t see, and only the events you care about need to be kept.
- Out-of-hours security & alarms
- Machine starts, stops & jams
- Counting & timing recurring events
Listening for what matters, without recording conversations.
Sound event detectionILLUSTRATIVE DEMO
glass_break 0.96reversing_alarm 0.91
- Background · wind & traffic
- Glass break — alert sent
- Vehicle reversing — logged
KEYHOLDER ALERTED · NO SPEECH RECORDED 0.5 sec
04 / NOISE & ENVIRONMENT MONITORING
Know what is making the noise.
A sound level meter tells you how loud it is; a model can also tell you what it is. Classifying noise by source helps sites show they are within limits, answer complaints with evidence and survey wildlife over weeks without hours of manual listening.
- Site, venue & event noise limits
- Investigating noise complaints
- Wildlife & bird surveys
Evidence by source, not just a decibel reading.
Noise & environment monitoringILLUSTRATIVE DEMO
0006121824
NOISE BY SOURCE
Plant & machinery46%
Road traffic38%
Wildlife & birds16%
ONE BREACH · SOURCE IDENTIFIED 24 h report
05 / VOICE COMMANDS & HANDS-FREE
Let busy hands work by voice.
Keyword spotting and speech recognition let staff log a reading, tick off a check or look up a job without putting down tools or taking off gloves. Small models can run on a phone or tablet on site, even where the connection is poor.
- Hands-free logging & checklists
- Picking, inspection & delivery notes
- Works offline on small devices
Records made in the moment, not remembered later.
Voice commands & hands-freeILLUSTRATIVE DEMO
Log temperature, fridge two, three point five degrees.
- action
- log_temperature
- location
- Fridge 2
- reading
- 3.5 °C
FRIDGE CHECKS · 11:42
Fridge 1 · 3.1 °C
Fridge 2 · 3.5 °C
Freezer
CHECK LOGGED WITHOUT PUTTING ANYTHING DOWN 0.6 sec
02 / AND MORE
Other places
audio can help.
Searching recordings
Find the moment a topic, product or phrase was mentioned across hours of calls or archived audio.
Translation & subtitles
Transcribe and translate speech for subtitles, training videos or customers who speak other languages.
End-of-line sound tests
Check each finished product sounds right — motors, fans, switches or speakers — as a quick quality test.
Cleaning up audio
Reduce background noise and level out volume so recordings are clearer before they are used or shared.
Animal welfare
Listen for coughing, distress calls or unusual activity in livestock sheds and flag changes early.
03 / GETTING STARTED
What a good project
needs to succeed.
Recordings from the real place
Audio captured where the system will run, with its usual background noise. A few examples of the sounds that matter go a long way at first.
A clear list of sounds
Agree which sounds or words matter and what should happen when one is detected, including how quickly.
The right microphone setup
Microphone type and placement make a big difference. Existing hardware can sometimes be used, or I can suggest a simple kit.
Privacy by design
Audio can be processed on the device and discarded, keeping only events or text, so conversations are not stored.
04 / YOUR SOUNDS, PUT TO WORK
Have something
worth listening to?
Tell me what you would like to hear, count or catch. We can talk through whether audio ML is a good fit and what a small first test could look like.
Open to projects in North Yorkshire and nationwide