Testing and fault-finding · Series, part 1

STI measures the channel, not the conversation

What a Speech Transmission Index figure tells a fire alarm technician, and the four things it never hears.

By Nigel Williams · Published 30 October 2026

For years, people repeated an odd bit of trivia: a duck's quack doesn't echo. In 2003, acoustician Trevor Cox and his team at the University of Salford tested it. It does. The quack is long and fades slowly, so its echo runs into it, and you hear one long quack instead of a quack and an echo.

The room had added something the ear couldn't separate from the sound. That smearing is exactly what the Speech Transmission Index (STI) measures.

If you manage a building rather than test its voice alarm, this is worth five minutes too. It explains what the STI column on your commissioning report can and can't tell you.

STI rates the journey, not the speech

STI compares a standard test signal at the start of its journey with what arrives at the listener. Everything in between counts, and together it's called the channel. Most of the damage is done by the room and the noise: reverberation smears the signal and noise buries it. The sound system plays a smaller part.

Think of a courier rating. It tells you the parcel arrived intact. It says nothing about whether the letter inside makes sense. STI rates the parcel.

The test signal stands in for one standard male talker. The current measurement standard, BS EN IEC 60268-16:2020, dropped the female version that earlier editions carried. The meter listens through one microphone, so it hears with one ear.

Speech is a loudness shape that rises and falls

Say "baboon". "Ba" starts loud and tails off, "boo" does much the same, and the "n" is quieter still. Inside that shape are faster rises and falls. The rises and falls are called modulations, and the gap between a peak and the dip after it is the modulation depth.

Now say "hippopotamus". Your lips and tongue move more, and faster. It has more fast modulations.

STI sorts the modulations by speed, 14 rates from 0.63 to 12.5 times a second, and checks each rate in seven frequency bands from 125 Hz to 8 kHz. Keep the two ideas apart. Modulation speed is how fast the loudness changes. The frequency band is how high or low the sound is.

Reverberation fills in the dips, fastest first

Picture your street on a summer's day. You can see the kerbs, the wipers on the cars, the ants on the path. Then it snows. A light fall covers the fine detail first. A heavy fall buries the cars as well.

Reverberation works the same way. Each sound hangs in the air and fills the dip before the next one arrives. A short reverberation time blurs only the fastest modulations, the "hippopotamus" detail. A long one blurs the slower ones too, until the syllables of "ba-boo-n" run together. It happens at every pitch, and STI checks every band.

Noise floods the dips to the waterline

Noise is more like a lake. Everything below the waterline disappears, and everything above it is untouched.

STI doesn't read the average level the way a sound level meter does. It reads what the noise does to the shape. The quiet parts of the speech go under, the dips fill up to the waterline, and the depth between peak and dip shrinks. STI turns that shrinkage into an effective signal-to-noise ratio, band by band.

A loud low sound hides a quiet high one at the same moment

Say "gazelle" slowly and listen to the "z". It's two sounds at once: a low buzz from your voice and a high hiss from your teeth. The hiss tells your listener it's a "z" and not an "l". Say it loudly and the buzz starts to swamp the hiss.

That's the upward spread of masking. It happens when the sounds overlap in time, and STI's calculation allows for it.

Turning it up helps against noise, and hardly at all against reverberation

Where noise is the problem, more level lifts the speech above the waterline and the STI rises. The masking then eats into each extra decibel, so the gains shrink as you go.

Where reverberation is the problem, turning it up does very little. The reverberant sound rises by exactly as much as the direct sound, so the snow is just as deep.

That's why the room and the noise decide most of the result. Keep reverberation and noise in check and almost any reasonably designed system will be intelligible. Most everyday spaces are fine. The ones that struggle are the difficult ones, with long reverberation or high noise, and there the fix starts with the room and the noise, not with more loudspeakers. A system can't shorten the reverberation. Whether more loudspeakers help depends on the room. In a reverberant room every extra one adds reverberant sound and another path to each listener. In a well-damped room, such as a library, sound falls away much as it does outdoors, so extra loudspeakers mainly bring the direct sound closer to each listener. You can try this yourself in the "Inside the STI number" section of our STI explorer.

Four things the number never hears

Because STI rates the channel, anything that isn't the channel is invisible to it. Listening tests show each of these can move real word scores by more than the gap between a pass and a fail.

The measurement standard also lists hearing-impaired listeners among its own known limitations.

So why do the standards rely on it?

A fair question. If STI can't hear the talker, why build commissioning around it? Because a talker isn't repeatable and a channel is. Within one talker, one language and a system that doesn't distort, STI tracks word scores closely: the method's authors reported about 5 % accuracy across 167 channels (Steeneken and Houtgast, 1980). That's its strength. It just answers a narrower question than the one people ask of it.

The Sydney platform shows the other side. On STIPA alone it would have failed. It passed on a properly run word-score test with real listeners.

What the number means on your sheet

An STI figure tells you how kind the room and the noise, and to a lesser degree the system, are to a standard test signal. It doesn't tell you how clear a real announcement will sound to a real person. The duck's quack was never the problem. The room was, and STI is very good at measuring rooms.

Next step. Next time you take a reading, write down who will make the real announcements, and whether any compression or processing is switched in. Then read the next article in this series: One STI reading is not a pass mark.