DEMO

Audio

A signal in the sound itself

In tests on two radio stations, Airplay Live found audio features statistically associated with how well songs held listeners. The association was already present in the songs' early plays.

The tested song's own listener response is not used to calculate its audio signal.

What does the model hear?
Inside the Sound →

Audio and Listener Hold

The tested song's listener response stays out. The audio signal still lines up with listener hold.

Audio signal vs. listener hold · plays 11–20

Dance UK

415 songs

Dance UK: audio signal and measured listener holdEach dot is one song. Large markers show five group averages with 95% confidence intervals. Higher values mean better listener hold. Values are standardized within the station.-6-40064Group 1: 83 songs; mean -0.27; 95% interval -0.49 to -0.06Group 2: 83 songs; mean 0.03; 95% interval -0.17 to 0.22Group 3: 83 songs; mean -0.01; 95% interval -0.21 to 0.20Group 4: 83 songs; mean 0.08; 95% interval -0.12 to 0.28Group 5: 83 songs; mean 0.17; 95% interval -0.08 to 0.40Audio signalHigher = associated with better holdMeasured listener hold

The Big 80s

300 songs

The Big 80s: audio signal and measured listener holdEach dot is one song. Large markers show five group averages with 95% confidence intervals. Higher values mean better listener hold. Values are standardized within the station.-6-40064Group 1: 60 songs; mean -0.02; 95% interval -0.24 to 0.21Group 2: 60 songs; mean -0.14; 95% interval -0.43 to 0.14Group 3: 60 songs; mean -0.22; 95% interval -0.51 to 0.06Group 4: 60 songs; mean 0.08; 95% interval -0.14 to 0.31Group 5: 60 songs; mean 0.30; 95% interval 0.07 to 0.51Audio signalHigher = associated with better holdMeasured listener hold
Each dot is one song. The larger markers show five group averages, from the lower to the higher end of the audio signal, with 95% confidence intervals. The wide scatter means the signal is not a reliable prediction for an individual song.
The charts use only songs whose audio came entirely from separate broadcast events (the Separate broadcast events control below).
Dance UK 415 songs · The Big 80s 300 songs · Standardized per station, zero = station average · Research measures, not POP Scores
Group averages as a table
Dance UK · standardized values
Audio groupSongsMean hold95% interval
183-0.27-0.49 to -0.06
2830.03-0.17 to 0.22
383-0.01-0.21 to 0.20
4830.08-0.12 to 0.28
5830.17-0.08 to 0.40
The Big 80s · standardized values
Audio groupSongsMean hold95% interval
160-0.02-0.24 to 0.21
260-0.14-0.43 to 0.14
360-0.22-0.51 to 0.06
4600.08-0.14 to 0.31
5600.300.07 to 0.51

The Signal Holds Up

A finding that only appears under one analysis is not a finding.

Control tests · each control reruns the analysis
AnalysisDance UKThe Big 80s
Original analysis Association detected783 songs · p = 0.010 Association detected403 songs · p = 0.010
Separate broadcast events Association detected415 songs · p = 0.010 Association detected300 songs · p = 0.010
Artists kept together Association detected783 songs · p = 0.010 Association detected403 songs · p = 0.010
Both controls together Association detected415 songs · p = 0.040 Not established300 songs · p = 0.090
The association survives each control separately on both stations. With both controls combined, it remains on Dance UK and stays inconclusive on The Big 80s.
Dance UK + The Big 80s · Listener response on plays 11–20 · Shuffled-data control runs

What Was Tested

Audio features were extracted from broadcast recordings, and a separate model was fitted for each station. Listener hold was measured from the station's online listener count, relative to normal background audience movement.

Each song's audio value came from a model trained on other songs. The control tests compare the results with shuffled listener-response data.

Five test groups were used, with ten different group assignments; the charts show the assignment closest to the median test result. Group intervals use song-level bootstrap resampling with the audio groups held fixed. They describe uncertainty in the group averages, not the full uncertainty of fitting the model. Control tests use 99 shuffled datasets; p = 0.010 is the smallest possible reported value. The detection threshold is p ≤ 0.05.

Scope and Limits

These are retrospective analyses of songs the stations kept playing, measured on each station's online-stream audience. They do not establish cause and effect, performance on another station, or predictions made before broadcast.

The listener data in these tests was collected at five-second intervals; current measurement runs at one-second resolution. Audio analysis is still research, not a product feature.

The listener-hold method has its own control tests.

Press release: Airplay Live Finds Radio Audience Hold Signal in the Audio Itself · October 12, 2026

For immediate release

Airplay Live Finds Radio Audience Hold Signal in the Audio Itself

HELSINKI, Finland, October 12, 2026 - In tests on two radio stations, Airplay Live found a signal in the audio itself that was statistically associated with how well songs held listeners. The association was already present in the songs' early plays.

The signal is derived from the audio without using data on how listeners have responded to the song being tested. The finding emerged as a byproduct of Airplay Live's research into how scores measuring songs' listener hold settle as plays accumulate.

Airplay Live, developed by Helsinki Underground Ltd, measures whether individual songs and talk topics hold a station's online audience or lose it, using the station's own live listener count rather than surveys or listener panels.

The listener count itself is nothing new. Stream servers have long reported how many people are listening at a given moment. Airplay Live now monitors that audience at one-second resolution alongside the station's now-playing data, comparing movement during each play with the station's normal movement at that time. Repeated plays build a station-specific score showing how reliably each song holds that audience.

The listener-hold method has been tested separately for repeatability, schedule bias and false signals, with the results and known limitations published at airplaylive.com/controls/.

The same measurement framework is applied to talk. Airplay Live transcribes spoken content and groups recurring topics and mentions, showing which subjects hold listeners and which lose them as evidence builds across broadcasts.

“We wanted to know whether a song actually keeps listeners on the station, not whether someone remembers liking it,” said Timo Salonen, founder of Helsinki Underground. “The station's own audience already answers that question every time a song plays. We just had to find a way to measure it.”

The audio finding adds a new layer to that work: measurable properties in the audio are associated with differences in listener hold. Airplay Live is continuing the research before deciding whether the audio analysis becomes a product feature. Potential applications range from radio music programming to music production.

Results are specific to each station measured.

Stations can request pilot access from Airplay Live. Setup requires a reliable concurrent listener count and a now-playing feed. There is no station-side installation, listener panel or encoding.

Charts and methodology for the audio finding are available at airplaylive.com/audio/.

Airplay Live is built and operated by Helsinki Underground Ltd in Helsinki, Finland.

Media contact:
Timo Salonen
Helsinki Underground Ltd
Tel. +358 400 852 524
timo@hkiug.com
airplaylive.com