The most dangerous failure is not an error message, but a confident-looking result
Whistles, humming, and electronic pure tones can all have clear periodicity, so a pitch detector may return stable values. If the rest of the pipeline checks only whether a fundamental frequency was detected, these sounds can enter primary-timbre, age-impression, or gender-tendency calculations and produce a well-formed report whose meaning is unsupported.
An audio system should not quietly equate "can be calculated" with "is appropriate to interpret." In CVoice, special-sound routing happens before standard timbre matching: the system first checks whether the audio contains enough natural speech structure, then chooses the appropriate result path.
Beyond pitch, observe how structure changes
Ordinary speech contains vowels, consonants, onsets, pauses, and changes in articulation. A steady whistle usually concentrates energy in a narrow band; a sustained pure tone varies even less; humming comes from the vocal folds but lacks the consonants and changing vocal-tract shapes of normal sentences. Individual metrics can overlap, so routing cannot rely on a single threshold.
CVoice combines measures such as spectral-energy concentration, bandwidth, the proportion of narrowband frames, spectral change over time, pitch movement, onset behavior, zero-crossing patterns, and short-term timbral change. The goal is not to identify "who" is speaking or decide whether a sound is synthetic, but to answer a narrower question: is this input suitable for a standard speech report?
How the product should respond after routing
| Input State | Handling | Misleading Conclusions to Avoid |
|---|---|---|
| Natural speech with sufficient quality | Proceed to the standard voice-impression flow | Still make clear that the result reflects only this recording's impression |
| Whistling, humming, or sustained pure-tone tendency | Proceed to a separate special-sound result | Do not generate standard age or gender impressions |
| Too short, too quiet, or heavily noisy | Suggest re-recording | Do not turn missing evidence into a label |
| Cannot be reliably distinguished | Conservatively reject or reduce interpretation strength | Do not give falsely definitive answers |
Special-sound results are excluded from primary-timbre and combination rankings. That removes some of the fun of appearing to "measure everything," but preserves the consistency of what the reports mean.
This gate is not the end of sound classification
Borderline cases still exist: melodic speech, breathy voice, vocal imitation, extremely short phrases, and heavily noise-reduced recordings may all contain both speech and non-speech features. The system can only make an engineering routing decision, not claim authoritative classification of every kind of vocalization.