Does audio quality affect transcription accuracy? Yes – and here's how much, by the numbers.

Before a transcription model hears a single word, one measurement has already narrowed how well it can possibly do: the signal-to-noise ratio, or SNR – the gap in decibels between the voice you want and the background you don't. One ASR provider's production-metrics analysis maps that gap straight onto accuracy, and the curve is steep. At 20 dB SNR (a quiet room, close mic) it reports a word error rate around 3.5%. At 15 dB, roughly 7%. At 10 dB, about 15%. At 5 dB, near 35%. At 0 dB, where voice and noise are equally loud, it passes 70%.

The important part is the shape. The penalty is not linear – it accelerates through the 2 to 14 dB band, which is exactly where that same analysis says most real-world recordings land. A few decibels of extra noise near the clean end cost you almost nothing; the same few decibels near the messy end can double your errors.

Those figures come from a single provider's synthesis rather than one controlled experiment, so treat them as the shape of the effect, not a universal constant. But the shape is what matters, and every study below traces the same outline: the recording, not the model, decides most of it.

Noise doesn't just muffle speech; it changes what the model does with the gaps. In a 2022 study in Frontiers in Signal Processing, an open-source engine (DeepSpeech) trained on clean speech scored a 12% word error rate on clean audio and 79% once the same speech was pushed through added noise and network distortion – nearly seven times worse, for words that were, to a human ear, still there.

There is a subtler failure hiding in that number. A 2026 preprint that stress-tested modern models reports that Whisper, at low SNR, makes among the fewest deletions but the most insertions: it fills uncertain stretches with plausible, confident, wrong text rather than leaving them blank. Bad audio, in other words, doesn't only produce gaps in your transcript. It can produce fluent sentences that were never said, which is worse, because they read as true. That preprint is not yet peer-reviewed, so take the mechanism as indicative.

If your recordings have several people talking at once, that is its own accuracy problem, and a common one – we cover it in how well AI handles crosstalk.

This is why a model's headline accuracy and the transcript you actually get can look like two different products. The same provider analysis puts a modern engine near 8.7% word error rate on controlled medical dictation – clean, single speaker, good mic – and above 50% on real multi-speaker clinical conversation, citing a clinical study for a 2.8 to 5.7 times degradation from benchmark to production. Nothing about the model changed. The audio did.

Benchmarks are recorded to be transcribed. Your meeting was not. When you read that a tool is "95% accurate," the honest question is: on what audio? We pull that thread in how accurate AI transcription really is and in our own accuracy benchmark.

Distance and hard walls add a second, quieter enemy: reverberation, the smeared echo of a voice bouncing around a live room. A 2026 benchmark took 1,600 clean LibriSpeech utterances and convolved each with real room impulse responses – the same words, digitally moved into echoey rooms – then measured the cost.

Whisper's large-v3 model took a reverb penalty of 2.31 percentage points. The tiny model took 15.50. Size buys resilience: a bigger model absorbs room acoustics that a small one cannot, but even the best model pays something for the echo. (This is a preprint, so read the exact points as directional.) The practical read: a carpeted, curtained, small room will out-transcribe a beautiful glass-walled conference room every time.

Move the microphone and you move the accuracy, sometimes off a cliff. A controlled corpus (SPEECON) recorded the same speakers at four distances at once, so distance was the only thing that changed. On a headset mic 2 cm from the mouth, digit recognition hit 95.22%. At a hands-free 10 cm, 92.44%. At a middle distance of 0.5 to 1 m, 89.54%. At a far-talk mic 3 m away, 61.18% – a third of the accuracy gone, from nothing but distance.

The research on messier speech agrees. In the CHiME-5 challenge, built from conversational dinner-party audio, the baseline system scored 47.9% word error rate on close-talk binaural mics and 81.3% on a distant Kinect array: a 33.4-point gap for the identical conversation, decided entirely by which microphone you read from.

Both of those results come from older HMM and DNN systems, so read the magnitudes as the direction and rough scale of the effect; a modern model is more resilient. But the cliff is still there, and the lesson is cheap to apply: get the mic close to the mouth. It beats almost anything you can do afterward.

Some quality you can fix in post. Bandwidth is not one of them – once the high frequencies are gone, they are gone. Telephone-band audio is sampled at 8 kHz, which discards everything above roughly 4 kHz, and that band carries the fricatives and sibilants (the f, s, sh, th sounds) that tell "fine" from "sign."

A Microsoft study measured the toll. The same voice-search system scored 27.47% word error rate on 16 kHz wideband audio and 53.51% on 8 kHz narrowband – roughly double the errors for the same speech. A model trained only on wideband degraded further still, from 29.96% to 71.23%, when handed narrowband input.

The rule that falls out is simple: keep your recordings at 16 kHz or better, and don't route them through anything telephone-grade. This is an older system on a small vocabulary, so modern numbers would be lower, but the 8-versus-16 kHz gap holds in kind. Bandwidth is one quality factor among several; accent is another, treated in transcribing accented English.

Compression is the one place the research offers reassurance. In the same SPEECON study, re-encoding close-talk audio as MP3 at 24 kbps – a file about a tenth of the original size – cost only about 3% accuracy with one feature set and 6% with another. Moderate compression is nearly free.

The collapse comes at the very bottom. At 8 kbps, accuracy fell from 95.22% to 21.02%. Below a threshold, the codec throws away the speech itself, and there is no model good enough to recover it. So the guidance is narrow and specific: don't fear a reasonable MP3 or AAC, but never hand a transcription tool an 8 kbps voice-note-of-a-voice-note. Light compression, or none, is the safe setting.

None of this means the model doesn't matter. Today's large, weakly-supervised models tolerate messy real-world audio far better than the supervised systems that came before them. Whisper's own paper reports a 55.2% average relative reduction in errors versus a strong supervised model on out-of-distribution audio, at near-identical clean-speech accuracy (around 2.7% on test-clean). That is real, and it is why a current engine is worth using.

But that is generalization headroom, not a rescue. The same body of research is clear that machines still need a higher SNR than human listeners to reach the same accuracy – bad audio remains bad audio, and the model overcomes less of it than we would like. Among models tested in one psychometric study, wav2vec 2.0 came closest to human performance; see the psychometrics of ASR. The order of operations is the whole lesson: the recording sets the ceiling, and the model works underneath it.

That is where Pepys sits, honestly. Pepys runs a current large speech model – the class of model shown here to handle messy audio best – so you start from a high floor. But every study above says the same thing: most of the win is upstream. Record close. Kill the background noise and the echo. Keep it at 16 kHz or better, lightly compressed or not at all. A modern model plus a quick review pass does the rest. This post is the why; the companion guide on improving transcription accuracy is the what-to-do. When your audio is clean, you can turn it into text and expect the numbers on this page to fall in your favor. For where AI transcription still hits its limits, see the limits in 2026.

Figures on this page are quoted from the studies linked below, several of them 2026 preprints; last verified 2026-07-13.