This appendix describes the signal stage of an anonymised language education workflow. The recorded project used a known beep marker; no customer audio or new detection run is included here.

Match the marker

Silence is a weak boundary. A speaker can pause inside an answer, background noise can fill a quiet interval, and recording levels can vary between files. The workflow had a stronger clue: a known beep placed between questions. Finding that marker first gave the rest of the pipeline a stable place to cut before it considered speakers, languages, or script alignment.

This is why the segmentation stage is described separately from voice activity detection. VAD can help identify speech, but it is answering a different question. The boundary detector needed to recognise a particular signal, not merely decide whether a frame sounded busy.

Turn peaks into boundaries

The legacy notes describe frequency analysis and cross-correlation with a reference template. The template represents the expected beep. As it slides across the recording, the correlation score rises when the local waveform resembles that template. Candidate peaks are then filtered and converted into intervals for downstream processing.

An illustrative version of that control flow looks like this:

template = known_marker()
for frame in audio:
    score = correlate(template, frame)
    if score passes the chosen threshold:
        keep the peak if it is separated from the previous peak
make segments from the ordered marker times

The pseudocode names decisions that still need configuration: the marker template, the threshold, the minimum separation, and how to handle a missing or repeated peak. Those values belong to the recording format. They should not be presented as universal defaults.

When deterministic matching helps

Template matching is a good fit when the signal is known and stable. It is easier to reason about than an open-ended classifier because the match has a concrete reference. The same property limits it. A changed beep, clipping, heavy noise, or a different recording chain can lower the score or create a false peak. A fallback based on energy or silence may still be useful for files without a reliable marker, but it answers a less specific question and needs separate review.

The output should retain the detected marker time and the surrounding segment boundaries. That makes a missed or shifted boundary visible in the browser editor. The later model stages can then work on bounded clips instead of trying to infer the whole recording structure at once.

Evidence boundary

The main case study reports 33 of 33 beep signals detected in one documented sample containing 28 segments. That sample result does not establish universal detection, and the public material does not reproduce the audio, environment, threshold, or provider settings. The applicable conclusion is narrower: when a recording format supplies a stable marker, deterministic signal matching can reduce the amount of boundary-finding left to the reviewer.