Why AI can't always recover speech from noise

Modern AI denoising is genuinely startling the first time you hear it: a voice lifted clean out of traffic noise that would have been unusable five years ago. Which raises a fair question: if it can do that, why does any recording stay broken? Why do the worst cases come back with words missing?

The number that decides everything: SNR

Signal-to-noise ratio is the loudness gap between the voice and everything underneath it, in decibels. A quiet home recording might sit at 35 dB SNR: voice comfortably above the noise. At 20 dB the noise is present but the voice dominates. At 10 dB they are close enough that untrained listeners start missing words. Around 0 dB the voice and the noise are equally loud.

AI models moved the usable line dramatically. Where classic spectral tools got uncomfortable below roughly 20 dB, a modern speech model produces clean, pleasant output from material at 10 and even 5 dB. That is the magic zone, and most real-world bad recordings live in it, which is why the tools feel unstoppable.

What happens at the floor

Here is the part the marketing usually skips. A denoiser can only amplify and reconstruct evidence that survived the recording. Loud syllables carry plenty of evidence through heavy noise. But speech is not uniformly loud: the ends of words, soft consonants, an unstressed “the”, all sit 15 to 20 dB below the stressed peaks of the same sentence. When the noise floor rises past those quiet components, their evidence is simply gone, drowned before it reached the microphone file.

At that point the model faces a choice with no good option: reconstruct its best statistical guess, or output the gap. Modern models are conservative on purpose (a guessed word is worse than a missing one), so what you hear at the limit is exactly what our extreme demo shows: a mostly-clean voice with the occasional soft syllable clipped away. The louder words made it; the quiet ones were never really in the file.

You can hear this line yourself

Our demo section includes a recording degraded to roughly 7 dB SNR with reverb on top, deliberately at the edge. The result is intelligible and clean, and it drops the occasional syllable, and we label it that way, because that boundary is a property of the physics, not a defect of one tool. Any vendor demoing perfect recovery from that territory is showing you either a milder file than claimed or a reconstruction that is partly fiction.

Practical consequences

• Record 10 dB safer than you think you need.The gap between “fixable” and “partially gone” is small, and you cannot tell which side you are on by ear while recording. Closer microphone, quieter room; both buy SNR for free.
• If a recording is at the edge, moderate settings twice beat maximum once. Aggressive single passes push the model into reconstruction territory sooner.
• Judge tools on edge cases, not best cases. Any denoiser looks perfect at 25 dB SNR. The honest comparison is what it does at 8, and whether the vendor will show you.