Two independent channels
The recording is not duplicated stereo. Its two channels contain substantially different audio.
Complete source recording
Waveform laboratory
This is the complete YouTube-derived recording used in the analysis. Click to seek. Drag across either waveform to create a selection. Drag the highlighted selection to move it, or drag its edges to resize it. A selection can be played once, looped, or enlarged with Zoom to selection.
Channel 1 — hold music/call side
Channel 2 — nearby microphone side
Signal routing
Channel isolation
Channel 2 contains the nearby microphone activity examined in this project. Isolating it sends that channel equally to both ears without attempting to invent or reconstruct missing sound.
Difference modes can suppress material shared by both channels,
but may also exaggerate codec and phase artifacts. They are
diagnostic views, not stronger evidence.
Loading complete recording…
Processed output
Live frequency analyzer
This analyzer displays the signal after the selected channel routing and filters. It is a listening aid, not a phonetic classifier.
Selected passage
Enter exact boundaries in seconds, or drag directly on the waveform.
Forensic listening controls All filters begin bypassed
Filtering can make weak material easier to inspect, but it
cannot restore discarded information. Extreme settings may
create or emphasize speech-like artifacts. Always compare
against the bypassed source.
Processing audit
Hear what the filters changed
Compare the routed signal before filtering, after filtering, and the sample-by-sample residue calculated as raw minus processed.
Residue is not recovered speech. It contains every change caused
by filtering, including attenuation, frequency-dependent phase
shift, compressor timing, and numerical difference. Treat it as a
quality-control signal, not as a separate speaker.
Measured from decoded samples
Selection analysis
| Channel | Peak | RMS | Crest factor | DC offset | Zero crossings |
|---|---|---|---|---|---|
| Channel 1 | — | — | — | — | — |
| Channel 2 | — | — | — | — | — |
Long selections are sampled at a regular interval to keep the page responsive. Values describe signal level and shape; they do not determine what words were spoken.
Analysis bookmarks:
Standard player and source download
Download browser MP3 · Download preserved YouTube WebM source
The waveform uses a constant-bitrate stereo MP3 generated from the preserved YouTube audio stream. Both channels remain separate.
Experimental model output
Machine-generated phonetic hypotheses
These are experimental machine-generated phonetic and word hypotheses—not
verified quotations, factual probabilities, or a certified forensic
transcript. Contextual recognizers can produce plausible language that
the audio does not support. Every candidate must be checked against raw
Channel 2 with processing bypassed.
No candidate phrase prompting
No name, phrase, hotword list, initial prompt, or expected sentence is supplied to the blind generation pass.
Scores remain separate
Acoustic, contextual, no-speech, and cross-run agreement values are not merged into a false “truth percentage.”
Alternatives are finite
Displayed choices are only those produced by the configured models and settings. They are not exhaustive.
Looking for the static hypothesis package…
Complete generation provenance and limitations
Summary
Hold music can be separated
During the opening, most hold music is concentrated on one channel while nearby microphone activity appears on the other.
No confirmed “Nolan”
No examined passage contains a stable and independently recognizable utterance of the name “Nolan.”
Source path
Claimed MDMR CXone release
→
YouTube upload
→
Highest-quality available Opus stream
→
Lossless analysis copy
→
Channel separation and filtering
The first step is based on the uploader’s provenance claim. This project has not independently obtained the native MDMR file.
Listen before reading
Candidate A
Listen to the raw recording first. Do not reveal the interpretation until after forming your own impression.
Raw isolated channel
Conservative cleanup
Consonant-focused diagnostic
This version emphasizes consonant energy and should not be treated as stronger evidence than the conservative version.
Current best interpretation:
“Shut the fuck up.”
The wording remains plausible across multiple versions, but the source quality does not justify calling it certain.
Listen before reading
Candidate C
Raw isolated channel
Conservative cleanup
Vowel/formant-focused diagnostic
This version emphasizes vowel and formant structure.
Current best interpretation:
“Oh my God,” followed by a possible “what did we do?” or
“what do we do?”
“Oh my God” is the more stable portion. The trailing words remain ambiguous, particularly the distinction between “did” and “do.”
What about the name “Nolan”?
We could not confirm a stable utterance of “Nolan.”
The name has a recognizable two-syllable structure. That structure did not remain consistently identifiable in the raw isolated channel or the conservative cleanup.
This does not prove that the name was never spoken. It means this YouTube-derived recording does not support confidently transcribing it in the examined passages.
Why captions matter
Ambiguous audio
Weak recordings contain incomplete phonetic information. Compression, noise reduction and distance can erase or distort important consonants.
Caption-induced priming
When suggested words are displayed before listening, the brain may fit ambiguous sounds to the written phrase. Once formed, that interpretation can become difficult to unhear.
Findings by confidence
| Finding | Assessment |
|---|---|
| Two substantially independent channels | High confidence |
| Hold music concentrated on Channel 1 | High confidence |
| Nearby microphone activity on Channel 2 | High confidence |
| “Shut the fuck up” | Probable |
| “Oh my God” | Moderate to strong |
| “What did/do we do?” | Possible |
| Use of the name “Nolan” | Not supported |
Method
- Preserve the downloaded YouTube stream unchanged.
- Create a lossless working copy.
- Examine codec, channels, duration and signal structure.
- Separate the two recorded channels.
- Review the raw microphone-side channel.
- Apply conservative, non-generative filtering.
- Compare interpretations across raw and processed versions.
- Listen before revealing proposed wording to reduce caption-induced priming.
No language model or generative speech reconstruction was used to
manufacture, complete or replace missing words.
Limitations
- The working source was transcoded through YouTube.
- The native MDMR file has not been obtained.
- Filtering cannot restore information discarded before upload.
- Aggressive processing may create or exaggerate speech-like artifacts.
- Repeated listening can increase expectation and pattern completion.
- Perceived wording does not independently establish conduct or intent.
Repository and reproducibility
The website source and selected public audio files are preserved in the public GitHub repository.
View website source View formal methods and quality controls