Home／Voice Knowledge／Result variation

Why the Same Person Gets Different Results Twice

"The same person" does not mean "the same signal." First identify which layer changed before deciding whether the cause is natural variation, input differences, or a software issue.

Author: Chione · Published / updated: 2026-08-18 · Interpreting results

First separate the three kinds of tests

Test methodReasonable expectationCommon sources of variation
Re-upload the same fileWith the same algorithm version, core features and base results should be consistent.Service version, catalog version, presentation-layer tweaks, or software defects.
Re-record the same sentenceThe overall direction may be similar while specific scores and rankings can vary.Intonation, distance, volume, pauses, state, and automatic device processing.
Record with a different device or environmentUseful only as a rough reference; differences should not be attributed directly to changes in the voice.Microphone frequency response, room reflections, noise reduction, encoding, and signal-to-noise ratio.

Why a small change can switch a label

A continuous acoustic space is often presented as discrete labels. Suppose "warm" and "clear" have nearly identical match distances: a 2% change in one spectral feature may swap first and second place. The label appears to change dramatically even though the underlying coordinates moved only slightly.

This is a threshold or ranking-boundary effect. A more honest interface should also show combined tendencies, proportions, or explanations instead of making the top label look absolute.

Determinism does not mean never changing

Determinism in CVoice Reborn means that the same normalized audio, algorithm version, and catalog version produce the same base result. CVoice also keeps display logic stable for the same file. But when the input waveform changes or the algorithm is upgraded, determinism does not require returning the old result.

If a service update changes feature normalization, fixes octave errors, or expands the character catalog, the date and version should be recorded. That is more reliable than preserving known errors for superficial consistency.

How to compare without misleading yourself

  1. Keep the same original file and first verify whether repeated uploads are stable.
  2. When re-recording, keep the device, room, distance, text, and time of day fixed.
  3. Record two or three takes for each condition and look for a common trend rather than choosing only the result you like best.
  4. Compare the eight continuous dimensions and quality indicators first, then look at the top label.
  5. Check before comparing across versionsPublic update log。
When an issue is worth reportingIf the same file repeatedly produces completely different core results within a short period on the same version, or the page shows unexplained large jumps, the cause is more likely to be a program, cache, or deployment inconsistency.

Next step

Read firstSix Types of Recording Variables That Affect Voice AnalysisEstablish a fixed procedure, then interpret the resultsReading ReportsUnderstand the hierarchy of labels and scores.