Why Your Recorded Voice Sounds Like Someone Else
The Voice Took a Different Route
Imagine replaying a voice message you have just sent. The words are unmistakably yours. The voice is less cooperative. It belongs to you, apparently, but it does not sound quite like the person who was speaking a moment ago.
The familiar explanation is that a recording leaves out the sound conducted through your head. That is a useful beginning. It is not a complete explanation of recognizing yourself—and it says remarkably little about whether you will like what you hear.
The distinction I want to keep is between three questions: How did this sound reach me? Is that my voice? Do I like it? A change in the first can affect the second without answering the third.
The speaker gets more than the microphone
When you speak, sound reaches your hearing through the air, but vibrations also travel through the tissues and bones of your head. Speaking supplies bodily sensation and motor signals as well as an airborne sound. Ordinary recorded playback through a loudspeaker does not reproduce that whole event. It offers an external sound without the same act of producing it.
That does not make the recording fake. It makes your live experience of speaking a poor reference for deciding whether the recording is a faithful account of the airborne signal. Nor is any particular phone-and-speaker combination an exact declaration of what everyone else hears. It is another listening situation, not a court of final appeal.
The bodily route is often reduced to extra bass. Actual measurements make that shorthand less tidy. In a 2010 study, Sabine Reinfeldt and colleagues estimated the relative bone- and air-conducted contributions during ten phonemes. Their author abstract reports the bone-conducted contribution dominating between 1 and 2 kHz for most of the phonemes, with differences between vocalizations. That is a particular threshold-related comparison, not a universal recipe for restoring your voice with a bass knob. I could access the complete abstract, not the full paper, so I cannot independently appraise its full methods or graphs here. [1]
More importantly, changing a spectrum is not the same as changing a decision about identity.
Half a second, partly somebody else
Pavo Orepic and colleagues made that decision unusually concrete in a 2023 experiment. They recorded people sustaining the vowel /a/, selected stable half-second clips, and used voice-morphing software to make blends between two voices. Some blends contained more of the participant's voice; others more of someone else's. The listener chose which person the blend resembled more. [2, §§2.2–2.3]
This is a deliberately stripped-down puzzle. No sentence, anecdote or characteristic turn of phrase gives the speaker away. Just a brief vowel, stretched across a synthetic border between identities.
The researchers presented the blends through air conduction and through a bone-conduction headset. If a listener's answers changed more sharply as the mixture moved from mostly other to mostly self, the resulting decision curve was steeper. That is the relevant improvement: decisions followed the blend more distinctly. It is not a measure of how pleasant or natural the voice sounded.
In the first study, with 16 participants, bone conduction sharpened that boundary when listeners had not just been given their unmorphed recorded voice as a reference. In later blocks, before each of which listeners heard that reference through the delivery mode being tested, the slope advantage was no longer statistically significant. The reference could have supplied an easier comparison strategy; because exposure always came in the later blocks, it was not a fully counterbalanced manipulation. The experiment does not prove that simply hearing your recording makes you comfortable with it. [2, §3.1]
Then came the useful control. What if bone delivery simply helped people separate any well-known voice from a stranger's? A second group of 16 judged blends of a familiar person's voice and an unfamiliar voice. The researchers did not find the same bone-conduction advantage. [2, §3.2]
A larger third study compared the tasks within one cohort. It recruited 52 people; nine failed a separate self-voice recognition control criterion and were excluded from the main blend analysis. In the remaining 43, bone conduction sharpened the self-versus-stranger and self-versus-familiar decision curves, but not the familiar-versus-stranger curve. These are three cohorts in one research program, not three independent laboratories reaching the same result. [2, §§2.1, 3.3, 3.5]
The especially good wrinkle is that the larger study's curves did not show a blanket urge to answer “me.” Part of the advantage was a clearer rejection of blends dominated by the other voice. Recognizing yourself can improve through a better not me. [2, fig. 3; §4.5]
The headset did not put the whole person back
It would be easy to turn this into a satisfying story: add the missing bodily vibration, restore the real voice, recognize yourself.
The paper itself gives reasons not to.
At the end of the larger study, listeners also judged unmorphed clips. In that simpler control, the researchers did not find a significant bone-versus-air advantage. The benefit was not universal even within their own experiments. That control came after the blend tasks, not at a person's first encounter with an ordinary voice message; its null result is not proof that the devices are exactly equivalent either. [2, §3.5]
And bone-conducted playback changes more than one thing. It changes how sound reaches the inner ear as well as introducing a bodily route. The authors favor an account involving accompanying vibrotactile input, but explicitly acknowledge that their experiment cannot cleanly isolate acoustic filtering from that input. In the larger study, the authors report loudness-matching at low intensities intended to avoid perceptible vibration; whether listeners could identify the source was inferred from reports, not formally measured. This was not a touch-only intervention. [2, §§2.3, 4.7]
The participants were listening, not talking. The motor signals of active speech were not tested. Nor were ordinary sentences, a lifetime of recorded messages or the feeling of cringing at your own voice. The endpoint was an identity judgment about a brief sound. No liking scale appears in this experiment. [2, §4.6; methods]
That leaves a better answer than “your skull adds bass.” The voice you know from speaking is embedded in a different physical event from the voice you receive in playback. Bone delivery can help distinguish self from other in a carefully defined task, but it is not a demonstrated restoration of the entire speaking experience. Unfamiliarity, misidentification and dislike are different outcomes.
The next experiment I would want is one that more tightly matches the sound while separating passive playback from active speaking. How much of the difference belongs to the route through the head, and how much to being the person making the sound?
The recording did not necessarily get you wrong. It left you listening from somewhere else.
Sources and reading limits
[1] Reinfeldt, S., Östli, P., Håkansson, B. & Stenfelt, S. (2010). “Hearing one's own voice during phoneme vocalization—transmission by air and bone conduction.” Journal of the Acoustical Society of America 128, 751–762. DOI: 10.1121/1.3458855. Complete author abstract accessed through Europe PMC; full text was unavailable in this study. Independent physical evidence, not replication of the recognition task.
[2] Orepic, P., Kannape, O. A., Faivre, N. & Blanke, O. (2023). “Bone conduction facilitates self-other voice discrimination.” Royal Society Open Science 10, 221561. Complete main text · DOI: 10.1098/rsos.221561. Main methods, results, discussion and figure captions read. No supplementary-file or raw-data analysis, playback or new listener experiment performed for this essay. The opening is an imagined scene.