It is an acoustic model that can inform AI systems; it does not inherently require machine learning.

Are formants the same as pitch?

No. Formants concern resonant spectral shaping, while fundamental frequency concerns repetition of a periodic source.

Can the source contain noise?

Yes. Turbulence is important in many speech sounds, and some sounds combine noise with voicing.

AI Glossary / Time series and signal processing / Advanced

Source-filter model

Definition

The source-filter model explains speech as an excitation signal shaped by the acoustic response of the vocal tract. For voiced sounds, vocal-fold vibration supplies a roughly periodic source, while the tract emphasizes some frequency regions more than others. It is a useful approximation for understanding and synthesizing speech, not a complete account of every interaction in the human voice.

What it means

The model separates two questions: what supplies the sound energy, and what shapes its spectrum? Changing the repetition rate of a voiced source changes harmonic spacing. Changing the vocal-tract response changes the pattern of spectral emphasis. This helps explain how someone can sing different notes while maintaining a recognizable vowel, or change vowels while holding approximately the same note.

How it works

For a short interval, a simplified linear, time-invariant model writes the speech waveform as the convolution of a source signal with a filter's impulse response. In the frequency domain, convolution becomes multiplication: Y(f) = E(f)H(f). Complex spectra include phase as well as magnitude.
Voiced excitation is often modeled using pulses; turbulent excitation can be represented by noise. Both may contribute to a sound. Resonances associated with the vocal tract are called formants. The model's filter may also incorporate other effects, depending on the chosen convention. Specify what is included before comparing parameters across implementations. Actual articulation changes over time, so a fixed filter is generally a local simplification.

Three practical examples

A speech-synthesis demonstration holds a filter fixed while changing pulse spacing. Listeners can explore a pitch change without deliberately changing the resonant pattern. This illustrates a controllable model, not perfect independence in real anatomy.
A speech-analysis tool estimates a spectral envelope separately from individual harmonic peaks. The distinction helps analysts avoid treating every visible peak as a separate vocal-tract resonance.
An educational voice interface switches between periodic and noise-like excitation. It demonstrates why a model containing only a periodic buzz cannot represent the full variety of speech sounds. These examples are possible uses, not results from a deployed product study.

Think of it this way

Think of an electronic instrument with a tone generator followed by an equalizer. One component supplies a sound; the other changes its spectral balance. A human vocal tract is more complicated, but the separation provides a practical way to reason about excitation and resonance without treating them as the same quantity.

Worked example

Construct an idealized periodic source with fundamental frequency 100 Hz. Consider only its first three harmonics, at 100, 200 and 300 Hz, with amplitude magnitudes 1, 0.5 and 0.25 in arbitrary units. Give a fictional filter magnitude gains of 0.5, 2 and 1 at those frequencies. The corresponding output magnitudes are 0.5, 1 and 0.25. The 200 Hz component becomes the strongest even though the source's fundamental remains 100 Hz.
This calculation does not mean that the fundamental must now be 200 Hz. Because the output still contains the 100 Hz component, this three-component construction retains a 0.01-second fundamental period. The strongest spectral component and the fundamental frequency are different concepts. Perceived pitch also requires more than simply selecting the largest magnitude.
Now retain the original source but change the three filter gains to 2, 0.5 and 1. Output magnitudes become 2, 0.25 and 0.25. Harmonic locations stay fixed while their relative strengths change. These are stipulated gain values, not measurements of a particular vowel or a complete anatomical filter.
For an independent time-domain demonstration, let the source samples be [1, 0, 1] and the filter impulse response be [1, 0.5]. With zero values outside both finite sequences, their linear convolution is [1, 0.5, 1, 0.5]. Each source pulse contributes a shifted copy of the filter response. Multiplying corresponding samples would instead miss that spreading operation. This short numerical filter is an arithmetic illustration, not a realistic vocal-tract model.

Illustrative case study

Fictional ClearVoice Lab creates a teaching display with separate controls for source repetition and filter shape. During testing, someone labels the largest spectral peak “pitch.” The team uses the constructed 100 Hz example above to show why that label can be misleading when a filter makes the second harmonic strongest.
It changes the display to distinguish fundamental frequency, harmonic amplitudes and filter response. It also labels its synthetic examples as simplified demonstrations and keeps listening comparisons separate from claims about naturalness. The scenario illustrates how the model can improve explanations; no recognition accuracy or commercial voice quality is inferred from the arithmetic.

Common misconceptions and mistakes

The filter does not create a new harmonic at every resonance in the simplified linear model. It changes the response to energy supplied by the excitation. Formants are not interchangeable with harmonics, and fundamental frequency is not simply the tallest spectral peak. Source and filter independence is a modeling assumption, not a claim that vocal-fold dynamics can never interact with acoustic loading. Unvoiced sounds also require attention to noise sources and where they arise.

Business use

Use this model to make speech-system requirements more precise: does a change concern pitch, spectral shape, excitation type or intelligibility? It can support controlled synthesis experiments and explanations of speech features. Validate the actual output with appropriate listening or task evaluations. A physically interpretable parameter does not by itself establish that users find a generated voice clear, natural or suitable for a service.

Check your understanding

Why does the first constructed filter produce an output amplitude of 1 at 200 Hz?

Answer: The source amplitude there is 0.5 and the filter magnitude gain is 2, so their product is 1. The result describes amplitude at that frequency, not a new source repetition rate.

Related terms

Fundamental frequency
Formant
Convolution
Linear predictive coding
harmonic
spectral envelope
impulse response
vocoder

Learn next

Change the short impulse response to [1, 0.25] while keeping source samples [1, 0, 1]. The linear convolution becomes [1, 0.25, 1, 0.25]. Explain why the second pulse contributes its own shifted response rather than modifying only the first pulse's output.

Convolution

Sources

John Coleman, University of Oxford, The Source-Filter Model of Speech Production: https://www.phon.ox.ac.uk/jcoleman/source_filter.htm University College London, Acoustics of Speech and Hearing, Lecture 7, Source-Filter Model: https://www.phon.ucl.ac.uk/courses/spsci/acoustics/lect1-7.pdf (indexed lecture text consulted; direct retrieval timed out). All numerical signals, gains and ClearVoice Lab are original teaching constructions.

Previous: Soundex Next: SOV language