Speech Quality vs. Speech Intelligibility

Clement Lauriat   Friday, 18th September 2026

Introduction

This post is about how we evaluate whether a spoken message can actually be understood by a listener, as opposed to how “good” or pleasant it sounds. In telecommunications, audio processing, and hearing science, two related but distinct concepts are often confused: speech quality and speech intelligibility.

A degraded signal can still be perfectly intelligible, and conversely, a technically clean signal can be hard to understand in some listening conditions. This article introduces the main subjective and objective methods used to measure intelligibility, from historical listening tests to modern automated algorithms, and gives an overview of where each method fits.

Speech quality and speech intelligibility

Since the end of the 90’s, ITU works on a way to qualify the quality induced by telecommunication systems through a metric, the Mean Opinion Score (MOS), a 5-point scale that translates the quality, or the effort required, to listen to a speech pattern.

While P.800 describes how to perform subjective tests, other methods were proposed by ITU to assess speech quality in an objective way: PAMS, PSQM, PESQ and the latest one, POLQA (P.863).

Some non-standardized objective methods were also proposed by labs like Google Lab, Berlin Telecom Institute, or communication solution builders.

Objective methods are designed to reach subjective quality of experience, which is why they have evolved over time as telecommunication technologies improved.

But speech quality is not intelligibility.

Where Speech Quality Assessment measures the fidelity of the audio signal along a path, Intelligibility answers the question: is the speech understandable?

We could say that if quality is good or excellent, intelligibility is expected to be good as well.

But in a noisy environment, or if some low-bit-rate codecs are used, quality may be low — the MOS may be low — without that meaning we cannot fully understand what is said.

This is where intelligibility metrics are useful. The next chapters introduce some subjective and objective concepts.

Subjective methods

Modified Rhyme Test (MRT)

The Modified Rhyme Test (MRT), developed by House et al. in 1965, is one of the most widely used subjective methods for assessing speech intelligibility at the phoneme level. It evaluates how well listeners can distinguish between words that differ by only a single phoneme, typically the initial or final consonant.

Principle: listeners are presented with a spoken word (transmitted through the system under test) and must choose, from a closed set of 6 rhyming words displayed on a card or screen, which one they heard. For example, for the target word “went”, the listener might choose among: went, bent, dent, tent, rent, sent.

Characteristics:

• Closed-set, forced-choice format: unlike open transcription tests, the listener selects from a limited list, which makes scoring objective and fast, but also means chance-level performance is non-zero (1/6 ≈ 16.7%).
• 50 word groups: the standard MRT uses 50 sets of 6 rhyming monosyllabic words (300 words total), covering a broad range of English consonants in initial and final position.
• Diagnostic value: because the confusions are constrained to single phonetic features, MRT results can highlight which sounds (e.g., voiced vs. unvoiced consonants, nasals, fricatives) are most affected by a given communication system or codec.
• Scoring: intelligibility is expressed as the percentage of correct identifications, corrected for guessing.

MRT has historically been used to evaluate telephone systems, military communication equipment, and low-bit-rate vocoders. Its main limitation is that it relies on isolated monosyllabic words rather than natural connected speech, so it may not fully capture the benefit listeners get from context (semantic, syntactic) in real conversations. This is precisely the gap that led to the development of automated variants such as ABC-MRT16 (section 4.a), which reuses the same rhyme-word structure but replaces the human listener with a signal-processing algorithm.

Other subjective methods

Besides the Modified Rhyme Test (MRT), several other subjective protocols have been used over the decades to assess intelligibility directly from human listeners:

• Diagnostic Rhyme Test (DRT): an earlier rhyme-based test, predecessor to the MRT, where listeners choose between two rhyming words differing by a single phonetic feature (e.g., voicing, nasality). It is still used in some military and telecom contexts (NATO STANAG).
• Phonetically Balanced (PB) Word Lists: listeners transcribe or repeat isolated monosyllabic words drawn from lists balanced to reflect the phoneme distribution of the language. Scoring is based on the percentage of words correctly identified.
• Sentence intelligibility tests (e.g., Harvard/IEEE sentences, SPIN test): use full sentences rather than isolated words, allowing evaluation of the contribution of semantic and syntactic context to intelligibility — closer to real-world listening conditions.
• CVC (Consonant-Vowel-Consonant) tests: focus specifically on consonant recognition, useful for isolating the impact of a system on particular phonetic classes.

These subjective methods remain the reference (“ground truth”) against which objective/automatic methods are calibrated, but they are costly, time-consuming, and require trained listener panels, which is why objective methods were developed.

MultiDSLA is an integrated hardware and software platform for the objective assessment of speech quality and intelligibility. It generates reference audio, injects it into the system under test, records the received signal, and applies the appropriate intrusive or non-intrusive algorithm. Supporting both analog and digital audio paths, MultiDSLA provides a complete, automated, and repeatable end-to-end testing workflow.