Signal-to-noise ratio (SNR)

Signal-to-noise ratio (SNR) is the ratio of the level of a desired signal to the level of background noise, expressed in decibels (dB). For speech, it measures how much louder the talker is than everything else in the recording, and it is among the strongest predictors of how accurately a system will transcribe the audio.

How signal-to-noise ratio works

Signal-to-noise ratio is defined as ten times the base-ten logarithm of signal power divided by noise power, which places it on a decibel scale. Higher numbers mean a cleaner signal: around 20 dB and above is comfortable for recognition, while single-digit SNR is genuinely hard for any model.

Because the scale is logarithmic, every 10 dB represents a tenfold change in the power ratio, so small dB differences matter more than they look.

SNR falls as the talker moves away from the microphone, as room noise rises, and as reverberation smears the sound, which means one voice can yield very different SNR depending on how it was captured.

Why signal-to-noise ratio matters for speech-to-text

Low SNR raises word error rate steeply, and the damage reaches beyond the recognizer. It confuses voice activity detection, which then struggles to tell speech from noise, and it increases the risk of a hallucination as the model tries to force meaning onto ambiguous audio.

Far-field audio captured across a room is the classic low-SNR case, which is exactly why techniques like noise suppression and beamforming exist in the first place.

How to improve signal-to-noise ratio

Capture audio as close to the source as you can, and clean it before recognition rather than hoping the model will absorb the noise. On difficult recordings, accuracy often tracks SNR at least as closely as it tracks your choice of provider, so improving the signal frequently buys more than switching models.

When you benchmark providers, include low-SNR samples that match your worst real-world conditions instead of testing only on clean studio audio.

Frequently asked questions

What is a good signal-to-noise ratio for speech recognition? Roughly 20 dB or higher is comfortable and gives models their best shot at low error rates. Below about 10 dB, accuracy degrades quickly no matter which recognizer you use.

How do you calculate signal-to-noise ratio? Take the ratio of signal power to noise power, then apply ten times its base-ten logarithm to get decibels. In practice you estimate noise power from silent stretches and signal power from speech.

How can you improve SNR before transcription? Move the microphone closer to the talker, reduce room noise and echo, and apply noise suppression or beamforming. Fixing capture conditions helps more than post-processing a bad recording.

Why does low SNR hurt transcription so much? Noise masks the frequency detail that separates one sound from another, so the recognizer guesses. It also throws off voice activity detection and can trigger hallucinated text on near-silent passages.