2026 · completed

Filtering noise and answering machines

Two small audio classifiers around the transcription pipeline: one filters out noise, the other recognises answering machines on outbound calls and updates the CRM.

ML research internship, AI Lab, Selectra, Madrid, June–September 2026.

precision of the noise detector
98.4%
answering machines correctly identified on a hard test set of 900 calls
95–97%
outbound calls checked per day
7,800

How it works

Top: a cascade keeps the noise detector cheap, since two statistical filters settle the obvious cases and only doubtful ones reach the heavier model. Bottom: the answering-machine detector reads only the first minute of an outbound call and writes to the CRM after the call.noisecall audiofilter 1statisticalfilter 2statisticalheavier modeldoubtful cases onlykeep or drop98.4% precisionobvious cases decided earlyoutbound calls, 7,800 a dayfirst 60 ssmall model95–97% correctCRM≈ 2 min after the callrecording deletedonce validatednever used for agents’ bonuses
Fig. 1 Top: a cascade keeps the noise detector cheap, since two statistical filters settle the obvious cases and only doubtful ones reach the heavier model. Bottom: the answering-machine detector reads only the first minute of an outbound call and writes to the CRM after the call.

Noise: two statistical filters settle the obvious cases, and only the doubtful ones reach a heavier model, so precision stays high at low cost. Answering machines: a small model listens to the first sixty seconds of each outbound call and writes the answer to the CRM about two minutes later; recordings are deleted once the result is validated.

Measured on the company’s own calls. I asked that the answering-machine flag never be used to compute agents’ bonuses.

Technical details

Noise filter

The transcription model sometimes writes words where there is only noise. The detector keeps this cheap: two statistical filters settle the obvious cases at almost no cost, and only the doubtful ones go to a heavier model. A large model on every segment would have cost far more for little gain.

Design
two statistical filters, then a heavier model on doubtful cases
Result
98.4% precision

Answering machines

An independent project: on outbound calls, find out automatically who answered, a person or a voicemail. A small model examines only the first sixty seconds of the call, and the result is written to the CRM about two minutes after the call. Recordings are deleted once the result is validated, and I asked that the flag never be used to compute agents’ bonuses.

Input
first 60 s of each outbound call
Volume
≈ 7,800 calls a day
Result
95–97% correct on a hard test set of 900 calls
Output
CRM updated ≈ 2 min after the call, audio deleted

← All projects