Speaking in noise

Why does a speaker in a noisy room reshape vowels rather than merely getting louder?

Walk into a loud bar and your voice changes before you have decided anything. It gets louder, obviously. But it also gets higher in pitch, slower, and — if you record it and look — the vowels themselves move: they become longer and more spread out from one another, and the energy in the sound shifts upward in frequency. Etienne Lombard described the loudness part of this in 1911, and the effect still carries his name.

The puzzle is why the extra machinery. If the problem is that you cannot be heard, the fix seems obvious: turn the volume up. Why should the shape of a vowel have anything to do with how much noise is in the room?

Begin with what "cannot be heard" actually means to a listener's ear. Their auditory system is not measuring your absolute loudness. It is trying to recover which sounds you made, from a mixture of your voice and the room. What matters is the ratio of your energy to the masking energy, and — crucially — it matters that ratio band by band across frequency, because the ear analyses sound in roughly independent frequency channels.

Now ask a question that decides the whole thing: is the noise in a crowded room flat across frequency? It is not. Room babble is other voices, and voices carry most of their energy low — in the region of the first few hundred hertz, where the fundamental and the first formant live. Add ventilation rumble and traffic and the low end is loud indeed, while the region above roughly two kilohertz is comparatively quiet.

So think about what a pure volume increase buys you. If you scale your whole spectrum up by ten decibels, you raise your energy in the bands the babble already dominates by the same amount you raise it in the quiet bands. Your ratio improves everywhere by ten decibels — which is real. But you have spent a great deal of vocal effort, and possibly hit the ceiling of what your larynx can produce, raising energy in bands that were losing badly anyway.

Now ask the other half: which bands actually carry the information? Vowel identity is carried largely by the first two formants — resonances of the vocal tract — and consonant place and manner cues live substantially higher still, in fricative noise and stop bursts above two kilohertz. Those upper regions are exactly where the masker is weakest.

That reframes the problem. You are not trying to be loud. You are trying to move your limited output power into the bands where it will survive. And that is what the measured changes do. Lombard speech shows a flatter spectral tilt — proportionally more energy at higher frequencies rather than a uniform lift. Vowel duration increases, which gives the listener's ear more time to integrate a weak signal. The vowel space expands, so vowels sit further apart from each other and a partially masked one is less easily confused with its neighbour. Fundamental frequency rises, which drags the harmonic structure — the thing that lets a listener group your voice as one stream separate from the babble — into a more distinctive region.

Is there evidence that this is genuinely better and not merely different? Yes, and it is the test that decides between "louder" and "reshaped". Take speech produced in noise and speech produced in quiet, and amplify the quiet speech to the same level. If loudness were the whole story, they should be equally intelligible. They are not: speech produced in noise wins, and studies by Lu and Cooke and others have found this gain repeatedly. The reshaping carries intelligibility that amplitude alone does not.

One honest complication. It is not settled how much of this is a listener-directed adjustment and how much is an automatic reflex of the speech motor system responding to its own disrupted auditory feedback. The effect appears even when there is no listener, which argues for something reflexive; but it is also modulated by whether you are addressing someone and by how much they seem to be struggling, which argues for a communicative component. Most current accounts allow both, with a reflexive core that a communicative goal can scale up or down.

The tempting summary — "you unconsciously optimise your speech for the channel" — overstates the precision. Nobody is computing band-by-band signal-to-noise ratios. What is happening is closer to a control loop that has been shaped, over a lifetime and probably over evolutionary time, to respond to degraded auditory feedback with a bundle of changes that happen to be useful together.

That bundle is also not free. Lombard speech is effortful; sustained it produces vocal fatigue and, in occupational settings, contributes to voice disorders in teachers and call-centre workers. And the reshaping can overshoot: the same expansion that separates vowels also makes speech sound strained, and speakers in noise routinely misjudge how loud they are, which is why a bar empties out and everyone is suddenly shouting.

One more refinement worth keeping. The effect scales with the noise, but not linearly and not without limit — the vocal apparatus has a ceiling, and past it the only remaining moves are the non-amplitude ones: slow down further, exaggerate further, and eventually give up on speech and gesture.

• How clear speech, produced deliberately for a hard-of-hearing listener, compares with Lombard speech produced reflexively.
• Why the effect is present in many other vocalising animals, and what that suggests about its origin.
• How speech recognition systems fail on Lombard speech when trained only on quiet recordings.

Every term the collection defines is gathered in the glossary.