Speech

All three pages in this section reduce speech intelligibility to a number in [0, 1], and the art is knowing which number answers your question. The Speech Transmission Index (STI) rates a transmission channel: a room, a public-address system, an intercom. The Speech Intelligibility Index (SII) rates a listening condition: this speech spectrum, in this noise, heard by this listener. A reverberant lecture hall is an STI problem; a hearing-aid fitting or a cockpit warning heard over engine noise is an SII problem.

The shared [0, 1] range is a coincidence of normalisation, not a common scale, and 0.6 means three different things on the three. An STI of 0.6 falls in band D of the IEC 60268-16 Annex F qualification ladder, whose eleven letters run from U below 0.36 to A+ at 0.76 and above; that is a good lecture room, and a voice-alarm specification typically sets its minimum a couple of bands lower. An SII of 0.6 means roughly 60 % of the importance-weighted speech spectrum is audible to that listener in that noise; the index is a fraction by construction and carries no standardised qualification ladder at all. A STOI of 0.6 has no absolute meaning: the mapping from the index to a percentage of words understood is fitted per listening-test corpus and is deliberately not implemented, so STOI is only ever read as a difference between two processors on the same material. Never substitute one index for another in a specification, and when a requirement quotes a number, check which standard it belongs to before computing anything.

The physical difference sits in what each index models. STI (IEC 60268-16) works on the speech envelope: intelligibility degrades when reverberation and noise flatten the slow intensity modulations of speech, and the index measures how much of that modulation survives the channel, via the modulation transfer function. It can be computed indirectly from a measured impulse response or measured directly with the STIPA test signal. Speech Transmission Index (STI) covers the modulation physics, both methods and the Annex F rating bands.

SII (ANSI S3.5-1997) works on audibility: intelligibility is predicted from how much of the speech-bearing spectrum rises above the listener’s effective threshold, band by band, weighted by each band’s importance to speech. Noise, self-masking, upward spread of masking and the listener’s own hearing threshold all enter explicitly, which is why SII extends naturally to hearing loss. Speech Intelligibility Index covers all four of the standard’s band procedures (critical band, equally-contributing critical band, one-third octave and octave), including the standard speech spectra for normal to shouted vocal effort.

A third pair of measures, STOI and ESTOI, answers yet another question: given a clean reference and a degraded or processed version of the same speech, how intelligible is the result? They rate the processing itself, which is why they are the standard yardstick for noise reduction and source separation.

The two standardised indices connect back to the rest of the library naturally: the STI consumes the impulse responses of Room Acoustics, and the SII consumes the hearing thresholds quantified in Hearing threshold. STOI and ESTOI have an upstream too, but a different kind of one: they take waveforms, so what feeds them is whatever produced the clean and the degraded recording — which is why they sit beside the signal-processing tools of Signals and spectra rather than beside a measurement standard.

Pages in this section

• Speech Transmission Index (STI): the IEC 60268-16 modulation transfer function, the indirect method from an impulse response, and direct STIPA measurement.
• Speech Intelligibility Index: the ANSI S3.5-1997 band-importance and band-audibility method, in noise and in hearing loss.
• Objective Intelligibility (STOI & ESTOI): the correlation-based measures for time-frequency weighted noisy speech, from a clean/degraded pair.

What this section does not cover

• No listener is tested, and no score is predicted. STOI returns the correlation index and not the percentage of words understood, because the logistic mapping is fitted per listening-test corpus; the SII returns an audibility fraction rather than a score; and no page here reproduces a subjective intelligibility test. No signal is acquired either: the STI page implements the STIPA direct signal and the indirect computation from an impulse response, but the full 14-modulation-frequency direct measurement of clause 6.3 is not implemented, so a chain with severe distortion needs measuring equipment rather than this library.

• Two coverage limits inside the SII are worth checking before use: the raised, loud and shouted speech spectra are carried for the one-third-octave procedure only, and the tabulated band-importance functions are each table’s average-speech compromise, with Annex B’s material-specific alternatives left to you through the band_importance= argument. There is no resampling between the four band procedures — each is fed spectra on its own bands. And the female speech option is not missing from the STI: Edition 5 of IEC 60268-16 removed it, so there is nothing left to implement.