Check the analysis window before naming an unwanted tone

Keep the recording and the analysis distinct

SciPy’s spectral-analysis tutorial explains that observing a finite interval produces spectral leakage. Windowing can reduce side lobes while widening the main lobe. A displayed shoulder therefore needs interpretation before it becomes a claim about another source.

Ask what question the audio owner is trying to answer. A suspected persistent hum, an intermittent click and an unfamiliar listening impression are not interchangeable investigations. Retain the unmodified recording and the exact segment under discussion. The reviewer should be able to distinguish a change in the input from a change in the way the input was examined.

Use a generated signal for the initial diagnostic because its intended contents are known. Keep this exercise separate from any actual interview or performance. A successful synthetic comparison can expose a misleading analysis habit without establishing what caused a real recording problem. Do not label a contributor, device or supplier as responsible on the strength of the demonstration.

Build a known-input comparison

Set sample rate to 8,192 Hz and retain 1,024 samples. Generate x[k] = sin(2π f k / 8192), for k from 0 through 1023, separately at f = 1000 and 1004 Hz. Apply NumPy rfft to each array and divide each magnitude array by its own maximum. These normalized displays do not report absolute level.

The unpadded grid is 8 Hz apart. At 1000 Hz, the first signal has its normalized peak of one; the adjacent 1008 Hz value is approximately numerical zero. For the 1004 Hz input, the 1000 and 1008 Hz values are 1.000000 and 0.996841. One generated tone can therefore occupy both displayed bins.

Record the complete settings beside both outputs. Do not compare a cropped segment with a longer segment while calling the inputs identical. Save enough information to reproduce the signal, not merely a screenshot with the frequency labels removed. The known-input case is useful precisely because the reviewer does not have to infer its intended contents from the plot.

Compare a taper and a denser display honestly

SciPy distinguishes a symmetric Hann window from its periodic spectral-analysis form. Here use hann(1024, sym=False), multiply the original samples by it, then transform. Preserve that choice explicitly; a window name without its construction is incomplete evidence.

For the 1004 Hz input, use transform length 8192 for both rectangular and Hann comparisons. Normalize each magnitude by its own maximum. At 1104 Hz the values are 0.024520 and 0.000164 respectively; at 1008 Hz they are 0.635615 and 0.848826. The selected outer value shrinks while the near-peak value increases. These are fixture measurements, not acceptance thresholds.

NumPy documents that requesting a longer transform pads the input with zeros. Transform length 8192 gives a 1 Hz grid here and a displayed maximum at 1004 Hz, but the original record still contains only 1024 acquired samples. The finer display has not collected another moment of sound.

Ask the owner to explain what changed before deciding which picture is more persuasive. A cleaner-looking result is not automatically a more complete account of the recording. Preserve both comparisons and the rationale for the chosen analysis. If the proposed conclusion depends on one unexplained setting, keep the interpretation open rather than approving a repair.

Write a bounded diagnostic conclusion

Separate the arithmetic finding from the production finding. This fixture demonstrates that one known input can produce different-looking displays under documented analysis choices. It does not identify an actual noise source, prove that a recording is acceptable or establish that listeners will notice a problem. A qualified sound reviewer must connect any real diagnosis to the intended delivery and listening evidence.

For the receiving workflow, request the original segment, analysis configuration, observed discrepancy and proposed next check together. If someone recommends processing the recording, ask what evidence would show that the intended content survived. Keep destructive editing outside this diagnostic exercise and retain the original source before any separately authorized production work.

After an analysis setting changes, repeat the known-input fixture and the relevant real-recording review separately. Record software versions and the scope of each check. The local execution used NumPy 2.3.5 and SciPy 1.17.0; the referenced manuals were checked on 7 November 2026. No player, listening session or production file was tested. The deliverable is an inspectable explanation of the display, not an automatic instruction to remove what it shows.

Make the reasoning concrete.

A fictional training-audio reviewer sees two strong neighboring bins and describes two unwanted sources. The generated single-tone comparison shows why that reading needs more evidence.

The reviewer corrects the unsupported interpretation and requests a bounded production investigation. No recording is altered.

Take the questions into your next brief.

Use these prompts in your own notes. Nothing is submitted from this page.

1. Question

   What exact production concern does the analysis address?

2. Input

   Which original segment and sample count were examined?

3. Configuration

   Can the complete transform and window choices be reproduced?

4. Comparison

   Which change belongs to the input and which to the display?

5. Conclusion

   What evidence remains before identifying a real source or proposing a repair?

Watch for the wrong conclusion.

• Calling every neighboring bin another source.
• Comparing normalized displays as if they reported absolute levels.
• Presenting additional transform points as newly recorded evidence.

Sources & scope

Original synthetic signal fixture and editorial review procedure. Primary references support the identified analysis behavior and its numerical applications. No listening assessment, real-source diagnosis, production processing or legal compliance claim is made.

• SciPy: Spectral Analysis →Primary software tutorial · Supports: Supports finite-window leakage and the windowing tradeoff.
• SciPy: hann →Primary software reference · Supports: Supports the periodic-window choice used in the numerical comparison.
• NumPy: rfft →Primary software reference · Supports: Supports real-input transforms and zero-padding behavior.