AI Glossary / Natural language and speech / Beginner
Pitch track
Definition
A pitch track is a time-indexed sequence of estimated pitch-related values, usually fundamental-frequency estimates from a sound recording. It shows how the estimated F0 changes during voiced regions and can mark intervals as unvoiced or uncertain. A plotted track is a representation of an analysis, not a complete copy of the sound or a direct transcription of its linguistic meaning.
What it means
The horizontal axis of a typical display represents time, and the vertical axis represents frequency in hertz or another specified pitch scale. Each point corresponds to an analysis time, often the center of a short window containing several waveform samples.
The track's time step is different from the audio sampling interval. Thousands of audio samples per second can support far fewer F0 estimates per second. A pitch track also differs from a spectrogram: it summarizes selected periodicity information instead of showing the full changing distribution of acoustic energy across frequencies.
How it works
A pitch-extraction method analyzes the sound, proposes candidate frequencies, and decides which estimates to retain across successive frames. Voicing decisions determine where the track contains an estimate and where it represents an unvoiced region.
Software can store multiple candidates and select a path through them, while a displayed contour usually shows the selected values. Joining plotted points can make the sequence look continuous, but interpolation or smoothing is an additional representation choice. It should not be confused with observing a reliable voiced frequency inside every gap.
Before interpreting the contour, inspect units, frame times, analysis range, and missing-value conventions. Sudden changes can reflect genuine acoustic behavior or extraction errors. A summary such as a mean also needs an explicit policy: selected voiced frames, interpolation, and time weighting can produce different quantities even when they begin from the same stored track.
Three practical examples
1. Prosody inspection: A researcher aligns a pitch track with words or syllables to examine where estimated F0 rises or falls within an utterance.
2. Synthesis review: A team compares generated contours with intended phrasing, using listening and other acoustic evidence to interpret suspicious sections.
3. Data-quality checks: An analyst flags implausible jumps or extensive missing estimates for review before using track summaries in a larger study.
Think of it this way
Think of a pitch track as a route drawn through a series of acoustic estimates over time. It helps reveal a trajectory, but the route can contain uncertain points and gaps. The analogy should not imply that the line traces a physical object or that every connecting segment has been independently measured. The stored estimates and the display policy must remain distinguishable.
Worked example
Consider an original hypothetical track with frame centers at 0.10, 0.12, 0.14, 0.16, and 0.18 seconds. The selected values are 120, 130, unvoiced, 150, and 160 hertz. The time step is 0.02 seconds, corresponding to 50 estimate positions per second.
There are five frames, but the interval between the first and last centers is four steps: 0.18−0.10=0.08 seconds. This does not specify the full audio duration because the analysis windows can extend on either side of their centers.
The mean of the four voiced estimates is (120+130+150+160)/4=140 hertz. Connecting the second and fourth voiced points with a drawn line would pass through 140 hertz at the missing center, but that interpolated value is not an observed voiced estimate. These independently checked numbers describe an invented track; no recording, interpolation experiment, or speech-tool output was analyzed.
Illustrative case study
Fictional case: Aspen Voice prepares charts for a review of generated announcements. One chart automatically connects points across unvoiced intervals, making the contour appear fully continuous. Reviewers initially mistake the connecting lines for measured values.
The team changes the display to mark gaps and documents which summary statistics use only voiced frames. It retains the original estimates so smoothing choices can be revisited. The scenario illustrates a communication problem in acoustic reporting, not a real experiment or measured product improvement.
Common misconceptions and mistakes
Do not read a smooth pitch track as proof of accurate extraction. An algorithm can consistently choose a wrong frequency candidate, including one an octave away. Conversely, a gap can represent a legitimate unvoiced region rather than a software failure.
Avoid confusing the number of frames with the number of intervals between their centers. Also distinguish analysis settings from display settings, and do not present interpolated points as original observations. Finally, the highest point in the curve is not automatically a pitch accent, and a rising line alone does not settle sentence meaning or speaker intention.
Business use
Pitch tracks make speech analysis reviewable when the application needs evidence about F0 behavior. They can help teams locate unusual intervals, compare analysis configurations, and communicate where a summary number came from.
For useful reporting, include units, time alignment, extraction method, and the treatment of unvoiced or missing data. Pair important acoustic claims with the relevant audio and task context when those are available. A chart should help the reviewer assess the evidence rather than hide uncertainty behind a continuous colored line.
People also ask
Is a pitch track the same as a waveform?
No. A waveform represents changing signal amplitude, while a pitch track summarizes estimated periodicity or F0 over analysis times.
Does every gap indicate silence?
No. Unvoiced speech, uncertain extraction, and other conditions can also produce missing voiced estimates.
Can I average every stored number?
Only after understanding the representation. Numeric placeholders for unvoiced or missing frames are not ordinary frequency observations.
Check your understanding
Why do five frame centers separated by 0.02 seconds span only 0.08 seconds from first to last?
Answer: Five centers create four intervening time steps, so the span is 4×0.02=0.08 seconds. Analysis-window extent is a separate matter.
Related terms
- Pitch extraction
- Fundamental frequency
- Spectrogram
- Pitch accent
- voicing
- frame
- time step
- interpolation
Learn next
Build a small table of times, frequencies, and voicing labels before plotting it. Compare a gap-preserving display with an interpolated one and explain which statements remain observations and which depend on a display or summary policy.
More in Natural language and speech
- Phrase structure grammar - A phrase structure grammar describes how smaller linguistic units combine into larger constituents, such as noun phrases, verb phrases, and sentences.
- Pipe - In regular expressions, the pipe symbol, |, expresses alternation: a pattern can match one alternative or another.
- Pitch - Pitch is the perceptual quality that makes a sound seem relatively high or low.
- Place of articulation - Place of articulation describes where the main constriction or contact occurs in the vocal tract when a speech sound is produced.
- Pointer-generator network - A pointer-generator network produces output by combining a vocabulary-based generation distribution with a distribution that copies tokens from the input.
- Possessive pronoun - A possessive pronoun is a pronominal expression that indicates a possession-like relationship without repeating a full possessor expression.
Sources
Sources: Praat manual, Pitch objects, frame times, candidates, and selected contours: https://www.fon.hum.uva.nl/praat/manual/Pitch.html . Praat configuration guidance, analysis versus display ranges: https://www.fon.hum.uva.nl/praat/manual/Intro_4_2__Configuring_the_pitch_contour.html . Jurafsky and Martin, Speech and Language Processing, Chapter 15, F0 and acoustic analysis: https://web.stanford.edu/~jurafsky/slp3/15.pdf . Relevant primary passages were read. The five-frame track and arithmetic are original and hypothetical; no audio was processed.
Previous: Pitch extraction Next: Pivot table