How ReCAP detects echo artifacts in satellite uplink chains

A satellite uplink rarely fails loudly. More often the picture holds and the sound arrives slightly wrong: the commentator's voice returns a fraction of a second late, so the broadcast mixes the live take with a smeared copy of itself. Listeners at home hear a presenter talking over themselves, and by the time the complaint reaches the truck outside the stadium, the artefact has already been archived, syndicated and cleared for reuse. This is the failure mode ReCAP's real-time analysis was built to catch, and it is a growing concern for Australian broadcasters running satellite news gathering across some of the most logistically demanding terrain in the world.

The project is an EU-funded research initiative for real-time content analysis and processing, working on automated metadata extraction, quality monitoring, face and logo recognition and duplicate detection across broadcast video. Echo in an audio return path sits neatly alongside those goals: it is a defect with a measurable signature, a clear timestamp, and an operational consequence. The technical question is how a detector separates a genuine acoustic echo from ordinary room reverberation, a compressor artefact or a fading carrier, and how it delivers an answer fast enough for a live gallery to act on.

Where echo artifacts come from in satellite uplinks

The audio path from the commentary box to the transmitter is long and loosely supervised. Microphones feed a sound desk, the desk feeds an encoder, the encoder modulates a carrier, the carrier climbs to a geostationary satellite, and the signal comes back down into an outside broadcast vehicle or a studio rack room where the director hears the return feed. Every junction in that chain is a candidate for a second copy of the signal to creep in. In Australia the physical conditions are unforgiving: a truck parked on gravel beside a stadium, rolled cables over hardstand, a folding microphone stand in a crosswind, and a generator running thirty metres from the talent.

Three families of cause dominate. Acoustic feedback happens when a talkback speaker sits close enough to an open microphone to re-enter it. Electrical feedback occurs when a reference feed, a talkback channel or a headphone monitor is patched into the mix instead of the monitor bus. Time-domain echo is the most stubborn of the three: the programme audio is returned from the studio or from a second truck and re-inserted into the mix with a fixed delay, so the echo survives the whole satellite round trip and arrives disguised as a delay line rather than as a fault. Weather adds to the confusion, since rain fade, a failing amplifier or a drifting clock can each bend the return path in ways that sound like echo to an untrained ear.

Symptom in the returned feed Likely physical cause Signature the detector isolates First operator action
Hollow, doubled commentary in the control room Acoustic coupling between microphone and talkback speaker Delayed self-copy with near-unity correlation above a few milliseconds Mute the offending monitor path, reposition the microphone
Nasal, band-limited voice Phase cancellation from a single reflection Spectral notches at multiples of the inverse path delay Check gain structure and mic placement on the desk
Periodic flutter or a faint "woof" under the mix Round trip through a second return path Comb filter spacing matched to the measured delay Re-align the return feed and confirm the delay setting
Echo present in the archive but absent live File-handling loopback during recording Flag attached to the stored asset, not to the live chain Quarantine the master and tag it for re-purposing
Programme audio consistently late against picture Clock or lip-sync offset Constant audio-to-video offset with no second correlation peak Hand the issue to playout engineering
Scratchy audio with momentary dropouts RF fade, rain fade or a faulty amplifier Short silent bursts correlated with carrier fade Pull the link margin and vehicle power reports

The last two rows matter because they look like echo and are not. A detector that cannot separate a room problem from a link problem will send crews chasing the wrong equipment, and in remote locations that mistake is expensive.

Measuring the return path and the echo signature

Detection starts with ground truth. To claim that a second copy of a voice exists, the system needs the delay between the direct signal and its replica, which means the measurement has to be tied to the actual signal path rather than to a fixed threshold. ReCAP's the NMR measurement framework pairs objective quality metrics with broadcast delivery targets, giving the audio analysis a common reference for what counts as a defect and how severe it is.

On a live uplink the measurements happen continuously. The analyser works on short windows of the returned audio, cross-correlates each window against the original programme feed, and looks for a second correlation peak that sits at a stable offset. When that peak is strong and persistent, the echo delay can be estimated to within a millisecond or two. From the delay the system derives the spacing of the comb filter the artefact will leave in the spectrum, which turns an abstract "double voice" into a number that can be checked against the truck's audio path configuration.

Severity is then expressed relative to programme loudness. Broadcast audio in Australia, as elsewhere, is mixed to a standardised loudness target, and the echo copy is graded as a fraction of the direct signal rather than as an absolute level. An echo buried well beneath the commentary draws a low-severity note, while a replica loud enough to be mistaken for a second speaker pulls the flag to the top of the scale and triggers an audible warning in the gallery. This matters in practice, because a truck parked at a cricket ground in Brisbane hears a stray voice over the PA all evening, and a system that treats every stray voice as an emergency gets ignored within a week.

What the echo detector actually hears

Three signal-level cues underpin the classification. In the time domain, an echo is a duplicated waveform, so the detector hunts for a correlation peak at a fixed offset and tests whether it persists across speech-active frames. A single anomalous frame might be a cough or a ducking artefact; a peak that reappears every few hundred milliseconds across a minute of commentary is structural. Persistence across changing talkers is the strongest evidence, because genuine room reverberation changes shape with the source while a delay-based echo preserves it.

In the frequency domain, the artefact betrays itself as comb filtering. A reflection at delay τ removes energy at regular intervals across the spectrum, and those notches can be detected even when the echo copy is too weak to stand out in the correlation trace. This catches shallow artefacts that a listener might tolerate but that accumulate into a thin, nasal quality across a whole segment of commentary. The spectral view also resolves overlapping echoes, where two reflections at different delays produce a more complicated but still characteristic pattern.

Real-time processing adds a third layer. A learned classifier, trained on labelled examples of echo against labelled examples of legitimate reverberation, settles the cases where the signal statistics are ambiguous. A cathedral-like reverb in a concert hall stretches and decays, while an echo repeats at constant amplitude. The classifier works on segmented audio, attaching a confidence score to each verdict, and it flags short windows rather than entire programmes so that a single contaminated line does not condemn a two-hour broadcast. For channels carrying music or effects rather than speech, the analyser switches to spectral and statistical cues, since waveform correlation has little to lock onto.

Turning an anomaly into a live flag

The output of the analysis is metadata, and metadata is what makes the flag useful. Each detection carries a timestamp, the channel identifier, the estimated echo delay, a severity score, a confidence value and the window of audio in question. That package travels into the quality monitoring layer, which keeps a rolling score per channel and highlights degraded periods on the monitoring dashboard. A director in an outside broadcast vehicle sees the fault labelled, hears the offending snippet on demand, and knows whether to fix the mix in the truck or escalate to the studio.

The same metadata is written into the recorded file as the programme is captured, which closes the loop between live operations and the archive. When a segment is later found, the flag sits alongside the asset's other attributes rather than living in a separate spreadsheet that nobody maintains. ReCAP's duplicate detection capability works on the same principle: a syndicated package appearing across several territories can be compared by content hash, and a defect detected in one copy can be traced into the others.

For Australian operations the payoff is straightforward. Regional feeds, multilingual outputs and syndicated sport packages all share source material, so an echo introduced at a single venue can reappear in dozens of downstream files. Flagging it once at the source, with a timestamp, is far cheaper than discovering it in a quality audit months later, after the master has been cleared for international sales.

Triage for engineers working across Australia

The first Australian reality is concentration. The ABC's operations sit at Ultimo in Sydney and SBS at Artarmon, and both run returns from vehicles and venues that feed several outputs at once. At SBS in particular, a defect in a shared source feed can contaminate dozens of language tracks simultaneously, so a flagged timestamp identifies the problem in a way a general "audio complaint" never would. Network Ten's Sydney and Melbourne facilities add to the traffic, and the afternoon check that crews call the arvo check becomes a natural point to review the day's flags and compare them against the truck's configuration.

The second is Melbourne, where a packed calendar of Australian rules football compresses the schedule into a few frantic weeks of the season and the Grand Final. Studios around Docklands and South Melbourne see multiple satellite uplinks live at once, and an echo that appears only during a sold-out stadium session is easily blamed on the venue acoustics when it actually originates in the return path. The third is Western Australia, where iron ore operations in the Pilbara run communication trailers that uplink to Perth. Those installations are exposed to heat, dust and long cable runs, and a talkback speaker clipped to a scaffold is a common source of acoustic coupling that would never survive a controlled studio environment.

The fourth is remote coverage such as the ABC's regional operations, where a studio in Alice Springs contributes bulletins that travel by satellite before travelling again over terrestrial distribution. Each additional hop multiplies the chance of a delay being introduced somewhere in the chain, and the operators at the far end cannot hear what the venue hears. Flagged anomalies with an estimated delay give them a specific thing to test. The practical routine is simple: read the flag, check whether the delay matches the truck's known return configuration, and if it does, mute or re-patch the path rather than assuming the satellite is at fault.

That routine has a security dimension, because these flags often reach the operator by email with a log attached. Alert mail impersonating an internal monitoring service is a well-worn trick, and a link or an attachment in one of those messages can be the actual problem instead of the echo. Crews are trained to confirm the alert through the dashboard rather than the message, to open attachments from the known internal source only, and to follow a short checklist on handling suspicious link clicks when something has already been opened. Reporting the message protects the rest of the crew, since the same impersonation tends to fan out across a network in minutes.

Echo metadata and the media asset library

Archived material inherits every defect it carries, so a flag attached at capture time remains valuable long after the broadcast. Media asset management teams can search for flagged assets, sort them by severity and pull the affected masters for repair before they reach a playlist, a partner feed or an international buyer. ReCAP's broader approach to archival object tagging is what makes this practical: the same tagging machinery that identifies objects in historical footage can carry an audio quality note on a current master, so one search interface covers both descriptive metadata and technical condition.

There is a second-order benefit. Once the archive knows that a specific segment was contaminated, the near-duplicate matching used for syndication can find every copy of that segment, including the ones still sitting in a regional station's local library. Australian broadcasters hold large volumes of this material, and cleaning it out by hand is close to impossible. A flagged asset, on the other hand, is easy to act on, and the difference shows up in what gets reused and what gets re-recorded.

Putting the flagging workflow into daily practice

The useful outcome of this work is a habit rather than a product. When an uplink engineer sets up a truck or a fixed site, the return path is measured once and the delay is recorded, so the system has a reference for what normal looks like on that particular path. When the analyser flags an echo, the response is defined before the fault happens: read the timestamp, compare the estimated delay with the recorded path, and re-patch or mute rather than troubleshoot blind. When a flag appears on an archived master, the asset is quarantined and searched for its duplicates before anyone schedules it again.

Applied consistently, that habit keeps a defect that would once have travelled unnoticed through a stadium feed, a regional studio and an international package now stays where it was created, with a timestamp, a severity score and a person responsible for fixing it.