Does formant synthesis require recorded speech units?

Classical parameter-driven formant synthesis can generate sound from an excitation model and filters, though recordings may inform parameter choices or analysis.

Can pitch and formants be adjusted separately?

Many source-filter synthesizers expose separate controls, while also allowing or modeling interactions between them.

Is every source-filter system a simple formant synthesizer?

No. Source-filter ideas support several representations and implementations, including more elaborate or learned systems.

AI Glossary / Natural language and speech / Intermediate

Formant synthesis

Definition

Formant synthesis is a method of generating speech-like sound by controlling an acoustic source and filters that model vocal-tract resonances. The synthesizer specifies formant frequencies, bandwidths, and other parameters over time, allowing the resulting waveform to be shaped without assembling it entirely from recorded speech segments.

What it means

Voiced speech can be modeled as an excitation source shaped by the vocal tract. Formant synthesis turns that idea into an explicit sound-generation process: create a source, define the resonance pattern, and combine them.
The parameters offer interpretable controls. Fundamental frequency affects the source’s repetition rate, while formant frequencies and bandwidths influence the spectral shape associated with vowel quality and other speech properties.
A complete speech synthesizer needs more than a few vowel settings. Timing, transitions, noise sources, amplitude changes, and consonant characteristics also matter. Formant synthesis describes an acoustic generation method; converting arbitrary written text into suitable parameter trajectories requires additional linguistic processing.

How it works

Choose an excitation model, such as a periodic source for voiced portions and noise components where appropriate. Specify the resonant filters and how their frequencies and bandwidths change during the sound.
Filters may be arranged in cascade or parallel structures, depending on the synthesizer. Additional components can model nasal resonances, spectral dips, frication, and interactions between source and filter.
Generate the waveform by applying the time-varying acoustic model. Then inspect and listen to the result, checking whether the intended contrast or utterance is perceptually clear.
The model is a controlled approximation. Explicit parameters are useful for experiments, but their presence does not guarantee natural speech or make every parameter combination a plausible human articulation.

Three practical examples

A phonetics experiment varies one formant while holding other parameters constant to study a perceptual contrast.
A speech-science lesson demonstrates how a common source can produce different vowel-like sounds through different filters.
A synthesis researcher explores controlled resonance trajectories while separately adjusting pitch, duration, and noise components.

Think of it this way

Think of a sound-making instrument whose excitation and resonating body can be adjusted separately. Changing the rhythm of excitation and changing the resonances produce different effects. The analogy is limited because speech involves many coordinated, time-varying processes, and an acoustic filter model is not a complete physical replica of a human speaker.

Worked example

Imagine a fictional 0.4-second voiced synthesis exercise. Set source F0 to 100 Hz, first-formant center F1 to 500 Hz, and second-formant center F2 to 1500 Hz. Keep the remaining required parameters fixed at explicitly chosen values.
For a second version, change only F0 to 125 Hz. Over 0.4 seconds, an ideal constant-rate source has 40 cycles in the first version and 50 in the second, because frequency multiplied by duration gives cycle count. The resonance centers remain at 500 and 1500 Hz.
For a third version, restore F0 to 100 Hz and instead make F2 rise linearly from 1500 to 1900 Hz during the 0.4 seconds. At the midpoint, 0.2 seconds, F2 is 1700 Hz. The source still has the original repetition rate.
These manipulations separate pitch-related source control from a changing filter resonance. They do not establish which vowel a listener will identify or how natural the sound will be. That requires the complete parameter configuration and perceptual evaluation.
The example is a checked parameter exercise; no audio was generated or listening test performed for this entry.

Illustrative case study

In a fictional project, Juniper Speech Lab prepares a classroom demonstration of source-filter theory. Students first hear two synthetic sounds with different source rates but unchanged resonance settings, then two with a changed resonance trajectory.
The instructor records every parameter setting so the comparison is reproducible and explains that the artificial sounds simplify real speech. Students discuss the perceived difference rather than treating a parameter label as a guaranteed perceptual outcome.
The imagined lesson illustrates controlled experimentation. It does not claim that a minimal formant model reproduces an individual voice or that listeners will identify all synthetic sounds in the same way.

Common misconceptions and mistakes

Formant synthesis is not simply playing a recording faster. Changing playback rate commonly changes several acoustic properties together, whereas a parameterized model can control selected properties separately.
It is not the same as concatenating recorded units, although hybrid systems can combine ideas.
Formant frequencies are also not interchangeable with F0. One describes resonance locations; the other describes a periodic source’s rate.
Finally, specifying formants alone does not create a complete text-to-speech system. Pronunciation, timing, prosody, and the source characteristics still need suitable treatment.

Business use

Formant synthesis is useful when interpretability and controlled acoustic manipulation are central goals, particularly in research, teaching, and specialized synthesis experiments. Teams can vary a selected parameter while documenting what was held constant.
For an application, evaluate intelligibility, naturalness, and whether the controls actually produce the desired perceptual differences. A technically valid waveform may still sound unnatural or fail to communicate the intended contrast.
Keep configurations and parameter conventions with the generated material. Frequency, bandwidth, amplitude, interpolation, and sampling choices all affect reproducibility, so a list of formant-center values by itself is not a complete synthesis recipe.

Check your understanding

A synthesizer keeps F0 constant but moves F2 upward over time. Has it necessarily increased the source’s pitch?

Answer: No. The specified change is to a filter resonance. It can change the spectral quality while the source repetition rate remains fixed, although perception depends on the complete sound.

Related terms

Learn next

Learn source-filter theory and the difference between harmonics and formants. Then inspect a synthesizer’s parameter definitions before making a controlled one-variable comparison. Progress from steady vowel-like sounds to time-varying transitions, and evaluate the audio rather than assuming the parameter values fully determine its perceived identity.

Sources

Praat, official source-filter synthesis tutorial: https://www.fon.hum.uva.nl/praat/manual/Source-filter_synthesis.html. Praat, filtering an excitation source with a FormantGrid: https://www.fon.hum.uva.nl/praat/manual/Source-filter_synthesis_2__Filtering_a_source.html. Praat, KlattGrid source, filter, and interaction documentation: https://www.fon.hum.uva.nl/praat/manual/KlattGrid.html. Independent review checked Praat’s FormantGrid filtering tutorial and KlattGrid source, filter, cascade/parallel and interaction definitions.

Previous: Formant Next: Formation matrix