PESQ, MOS and SNR: how to use each to evaluate audio models

Jonas Petrak, Research writer (AI-assisted) · · updated

When you need to judge how well an audio model performs, the right metric depends on what counts as "good" for your use case. If you care about how people actually hear the result, Mean Opinion Score (MOS) is the standard. If you want an automated, reproducible test for speech, Perceptual Evaluation of Speech Quality (PESQ) is widely used. For measuring how much noise remains, Signal-to-Noise Ratio (SNR) gives a clear, technical answer. Each metric answers a different question, and picking the wrong one can hide real problems or inflate weak results.

MOS captures what listeners actually hear

The Mean Opinion Score (MOS) is the go-to for measuring audio quality as experienced by real people. In a typical MOS test, you play audio samples for a group of listeners in a controlled setting—think a quiet room, with volume and background noise tightly regulated—and ask them to rate the quality on a scale, usually from 1 (bad) to 5 (excellent) [3]. The MOS is simply the average of all these ratings.

This method is powerful because it reflects actual human perception, not just technical correctness. For example, a speech enhancement model that removes noise but leaves speech sounding robotic might still score poorly on MOS, even if its technical metrics look good.

But MOS has tradeoffs:

  • Cost and time: You need to recruit listeners and run the tests, which can be expensive and slow.
  • Context sensitivity: MOS scores depend on how the test is run—room acoustics, listener expectations, and even language can shift the results [3].
  • Comparability: Scores from different tests or labs can’t be compared directly unless the experiments are carefully matched.

If you want to know how your model sounds to people, MOS is the gold standard. But if you need quick, repeatable, or large-scale testing, you’ll need something else.

PESQ gives an objective, repeatable score for speech

Perceptual Evaluation of Speech Quality (PESQ) is the industry standard for automated speech quality assessment [1]. PESQ compares the model’s output (the “degraded” signal) to a clean reference recording, analyzing them sample by sample after aligning them in time. It then predicts a quality score that closely matches what human listeners would report on the MOS scale.

PESQ is widely used by telecom companies, codec developers, and network testers because:

  • It’s objective: You get the same result every time for the same input.
  • It’s fast: No need for human listeners.
  • It’s specialized: It’s designed for speech, not music or general audio.

However, PESQ has limits:

  • Full reference required: You must have a clean reference signal for every test sample.
  • Speech-only: It models how humans hear speech, so it doesn’t work well for music or environmental sounds.
  • Licensing: Using PESQ for commercial purposes requires a license.

A telecom engineer evaluating a new voice codec, for example, would use PESQ to benchmark quality across different network conditions. But if you’re working on music enhancement or don’t have reference signals, PESQ won’t fit.

SNR measures how much noise is left

Signal-to-Noise Ratio (SNR) is a classic metric that tells you how much of the output is useful signal versus unwanted noise [2]. It’s calculated as the ratio of signal power to noise power, usually expressed in decibels (dB). A higher SNR means a cleaner signal.

For example, if your model outputs a speech signal with a power of 1 unit and the noise power is 0.1 units, the SNR is 10, which works out to 10 dB. This makes it easy to compare models or track improvements as you tweak parameters.

SNR is:

  • Simple and universal: Works for any audio, not just speech.
  • Objective: No human raters needed.
  • Direct: Shows how well noise is suppressed.

But SNR has its own caveats:

  • Doesn’t capture artifacts: A model could remove noise but introduce distortions that SNR won’t catch.
  • Reference required: You need to know what the “clean” signal should be.
  • Not always aligned with perception: A signal with high SNR can still sound bad if artifacts are present.

SNR is best when you care about technical clarity—say, in a hearing aid algorithm or a denoising tool for recordings. But it won’t tell you if the result actually sounds good to listeners.

A worked example: Comparing MOS, PESQ, and SNR

Suppose you’re building a speech enhancement model for mobile calls and want to assess its performance under typical street noise.

  • MOS test: You invite 20 listeners to rate the enhanced audio in a controlled room, using a 1–5 scale. The mean score is 4.2. This puts your output solidly in the “good” to “excellent” range as perceived by listeners—useful for flagging artifacts or distortions that objective metrics might miss.
  • PESQ score: Running the same processed audio through the PESQ algorithm (which compares it against the original clean reference), you get a score of 4.0 (on the MOS scale). This aligns closely with the human MOS ratings, suggesting your model’s perceived quality is reliably captured by PESQ for this speech sample.
  • SNR calculation: You measure the output’s signal and residual noise powers: if the signal is 1 and noise is 0.1, your SNR is 10 (10 dB). This quantifies the noise suppression: the model reduces noise power by a factor of 10, but SNR alone won’t flag if the speech sounds robotic or unnatural to listeners.

In practice:

  • MOS reveals if listeners are satisfied—a direct check on perceptual quality.
  • PESQ lets you automate quality checks during model iteration, as long as your domain is speech and you have a clean reference.
  • SNR quickly tells you about noise reduction, but can’t surface listening fatigue or subtle artifacts.

Relying on SNR alone can hide perceptual issues; relying only on MOS means manual re-testing every time you tweak parameters. PESQ provides a middle ground for speech, but its reliability drops for music or mixed audio. For the most robust evaluation, run all three and compare where they diverge.

Method: MOS and PESQ as described in ITU-T P.800 and P.862; SNR from standard signal/noise measurement.

The tools that do this, and how well

  • MOS: Run by human raters using standardized protocols (e.g., ITU-T P.800), sometimes estimated by AI models trained on human scores [3].
  • PESQ: Implemented in commercial and research tools (requires license); used by telecoms and codec developers for speech [1].
  • SNR: Calculated in any audio analysis tool or script; used in audio engineering, hearing aid development, and denoising research [2].

For speech models, PESQ and MOS often track closely, but only MOS captures real-world listener reactions—including negative ones to subtle artifacts. SNR is easy to compute and track, but should not be your only metric if you care about subjective quality.

Questions readers ask

What if I don’t have a clean reference signal?

You can’t use PESQ or SNR directly without a reference. In these cases, use no-reference metrics (like ITU-T P.563 for speech), but expect lower accuracy. For subjective quality, MOS tests with human listeners are still possible.

Can I use PESQ for music or environmental sounds?

No—PESQ is tuned for speech. For music or general audio, look into other perceptual metrics like PEAQ or run MOS tests with listeners.

How do I know if a MOS difference is meaningful?

MOS is collected as an average rating from human listeners, typically on a 1–5 scale. The rating reflects ordinal categories—Bad to Excellent—not precise intervals, so the size of a difference should always be reported alongside the context and setup. Larger gaps are easier to interpret, but the meaningfulness depends on your test design and what listeners actually report [3].

How the numbers work

Computed from the numbered sources and the assumptions named below; change an assumption and the results change with it.

FigureValueWhere it comes from
Minimum MOS score1 MOS pointssource [3]
Maximum MOS score5 MOS pointssource [3]
Reference signal power1 arbitrary unitsgiven: SNR is a ratio, so use normalized signal power
Noise power0.1 arbitrary unitsassumption: Example SNR with 10x signal over noise
MOS score range4 MOS pointscomputed: Maximum MOS score - Minimum MOS score
Signal-to-noise power ratio10 ratiocomputed: Reference signal power ÷ Noise power
Signal-to-noise ratio (decibels)10 dBcomputed: 10 × log10(Signal-to-noise power ratio)
Work it out with your own numbers
  • Minimum MOS score: 1 MOS points
  • Maximum MOS score: 5 MOS points
  • Reference signal power: 1 arbitrary units
  • Noise power: 0.1 arbitrary units

Works out: MOS score range; Signal-to-noise power ratio; Signal-to-noise ratio (decibels)

*Corrected 2026-11-21: The title now answers the question first. Headline was: "Use MOS, PESQ, and SNR to evaluate audio models".*

Sources

  1. Perceptual Evaluation of Speech Quality - Wikipedia
  2. Signal-to-noise ratio - Wikipedia
  3. Mean opinion score - Wikipedia

More articles