Exploratory acoustic feature extraction, quality control, within-subject reference estimation, and longitudinal analysis for investigational voice research.
Version 1.0.0 · Last updated August 2026
Cogitrac Voice is an investigational research platform designed to extract acoustic and spectral measurements from voice recordings for longitudinal research. The platform supports exploratory investigation of whether voice-derived variables may be associated with within-subject changes or prospectively defined study outcomes.
Investigational research platform. The extracted measurements are candidate research variables. They have not been established as validated biomarkers of epilepsy, seizures, neurological disease, or any other clinical condition. Individual measurements should not be interpreted as diagnostic, prognostic, treatment-related, or clinically actionable results.
Feature extraction is intended to preserve potentially informative acoustic measurements for subsequent research. The fact that a feature is extracted or stored does not establish that it has clinical significance, predictive value, or a causal relationship with a study outcome.
| Component | Status | Interpretation |
|---|---|---|
| Mean F0 | Extracted | Candidate acoustic variable for research analysis. |
| F0 variability | Extracted | Candidate within-recording variability measure. |
| Jitter / shimmer | Conditional | Exploratory voice-quality measurements requiring sufficient detected pulses. |
| Formants | Extracted | Candidate spectral/articulatory measurements subject to recording and algorithm limitations. |
| MFCCs | Extracted | Candidate spectral representation for exploratory analysis. |
| Custom derived features | Derived | Exploratory variables requiring independent evaluation and validation. |
| Seizure detection | Not validated | No clinical seizure-detection claim is established by this methodology. |
| Clinical biomarker threshold | Not established | Requires appropriately designed development and independent validation studies. |
Voice measurements can be affected by recording equipment, device processing, microphone position, background noise, room acoustics, speaking task, language, hydration, fatigue, illness, medication, emotional state, and ordinary biological variation.
Consequently, longitudinal comparisons should be interpreted in the context of recording conditions and relevant protocol information.
Research consideration: A change in an acoustic measurement may reflect recording conditions, ordinary biological variation, task differences, or another confounding factor rather than a change in the study outcome of interest.
Recordings undergo basic quality checks before acoustic feature extraction. The current extraction implementation includes an overall signal-level check intended to identify recordings that are effectively silent or extremely low in amplitude.
PointProcess contains at least three detected pulses.
np.nan and should be excluded or otherwise handled
according to the downstream analysis plan.
1e-5 in the current implementation.
Implementation limitation: The extraction code described here does not, by itself, establish a universal minimum three-second recording duration or perform general-purpose background-noise removal. If recording eligibility, denoising, silence trimming, normalization, or duration requirements are implemented elsewhere in the production system, those procedures should be documented separately and versioned.
Reproducibility: RMS thresholds are dependent on the amplitude representation of the supplied waveform. Exact reproduction therefore requires documentation of audio scaling, normalization, resampling, channel handling, and any upstream signal processing.
Feature extraction uses Parselmouth, a Python interface to Praat, together with Librosa for complementary digital signal-processing features.
The current pipeline extracts a broad set of acoustic and spectral variables. These measurements are retained as candidate variables for exploratory longitudinal analysis, reproducibility studies, feature selection, and future research.
Interpretation: Extraction does not imply validation. A feature can be scientifically useful for hypothesis generation without having established clinical, diagnostic, prognostic, or neurological significance.
Mean fundamental frequency across detected voiced pitch frames.
Parselmouth · ExploratoryStandard deviation of detected voiced fundamental-frequency values.
Parselmouth · ExploratoryCorrelation between two segments of the detected F0 sequence, used as an exploratory measure of within-recording F0 consistency.
Derived feature · ExploratoryLocal cycle-to-cycle variation in detected period duration, calculated using the configured Praat local-jitter procedure.
Parselmouth / Praat · ExploratoryLocal cycle-to-cycle variation in detected period amplitude, calculated from the Sound and PointProcess representations.
Parselmouth / Praat · ExploratoryMean harmonicity-derived harmonic-to-noise ratio across the analyzed recording.
Parselmouth / Praat · ExploratoryMeasurement caveat: Jitter and shimmer are established acoustic voice-quality measurements, but their reliability depends on recording characteristics, pitch tracking, voice type, task, and analysis parameters. Their use in application recordings should therefore be considered exploratory unless reliability has been demonstrated for the specific recording protocol.
First-formant estimate obtained at the midpoint of the analyzed recording when a valid formant estimate is available.
Parselmouth · ExploratorySecond-formant estimate obtained at the midpoint of the analyzed recording when a valid formant estimate is available.
Parselmouth · ExploratoryThird-formant estimate obtained at the midpoint of the analyzed recording when a valid formant estimate is available.
Parselmouth · ExploratoryCustom derived measure based on variability in sampled F1 and F2 values across the recording. It is exploratory and has not been independently validated as a measure of articulatory instability.
Derived feature · ExploratoryFormant eligibility: A minimum recording duration should not be inferred from this extraction description alone. If the study protocol requires a minimum duration, that requirement should be specified separately as a recording-eligibility criterion and versioned with the study protocol.
Mean root-mean-square energy across short-time frames.
Librosa · Extracted / retainedMean spectral centroid across short-time frames.
Librosa · Extracted / retainedMean spectral bandwidth across short-time frames.
Librosa · Extracted / retainedMean spectral rolloff frequency across short-time frames.
Librosa · Extracted / retainedMean zero-crossing rate across short-time frames.
Librosa · Extracted / retainedCorrelation between mean spectral magnitudes in the low-frequency (<1 kHz) and mid-frequency (1–4 kHz) bands.
Derived feature · ExploratoryThirteen MFCCs calculated over short-time frames and summarized as recording-level means.
Librosa · Extracted / retainedFeature status: The extraction layer is intentionally broader than any individual research hypothesis. Features may be retained for later analysis, feature-selection studies, replication, and hypothesis generation. Inclusion in the extraction pipeline should not be interpreted as evidence of association with epilepsy, seizures, or another neurological outcome.
Where longitudinal analysis is performed, measurements may be compared with prior recordings from the same participant. The purpose of a within-subject reference is to characterize observed individual variation rather than assume that a population-level reference value is clinically meaningful for every participant.
Study-specific reference and event windows should be defined prospectively in the research protocol. For example, a study may designate recordings collected within a specified period before an independently documented event as event-proximal observations while using separate recordings for baseline estimation.
Important: Reference-window definitions are study-design decisions, not validated clinical rules. Where feasible, they should be specified before outcome analysis to reduce analytical flexibility and potential bias.
The number of recordings required for baseline estimation should be determined by the study design, expected within-subject variability, statistical model, desired precision, and available data.
Approximately 30 recordings may be a practical starting point for some exploratory longitudinal designs, but this value is a design heuristic, not a validated biomarker threshold or universal minimum.
Feature extraction and statistical analysis are separate stages of the research pipeline. The extraction layer produces acoustic measurements; downstream analyses may evaluate distributions, within-subject variability, temporal patterns, associations with study outcomes, and candidate feature performance.
For exploratory analyses using within-subject reference standardization, a conventional standardized score may be calculated as:
Z = (X − μref) / σref
where X is the current measurement, and μref and σref are the mean and standard deviation estimated from the selected within-subject reference period.
A standardized score describes deviation from the selected reference distribution. It is not, by itself, a biomarker, diagnostic threshold, seizure probability, or measure of disease severity.
Research pipelines may use predefined statistical thresholds to identify observations that are unusual relative to an individual's reference distribution. Under an approximately normal reference distribution, ±2.58 standard deviations corresponds approximately to a two-sided 99% reference interval.
Statistical caveat: This correspondence assumes an approximately normal reference distribution and does not account for uncertainty in estimating the participant-specific mean and standard deviation. It also does not establish that an observation outside the interval has clinical significance.
Research interpretation: A value outside a predefined statistical threshold indicates an unusual observation relative to the selected reference distribution. It does not establish that the observation is seizure-related, neurologically significant, or clinically actionable.
Longitudinal analyses should account, as appropriate, for repeated measurements within participants, temporal dependence, baseline drift, missing observations, multiple comparisons, recording-condition effects, task differences, and relevant participant-level covariates.
Statistical thresholds and model specifications intended for confirmatory analyses should be defined prospectively where feasible. Exploratory findings should be clearly distinguished from confirmatory results.
Extracted features may be evaluated as candidate research variables based on statistical behavior, measurement reliability, reproducibility, association with prospectively defined study outcomes, temporal relationships, and robustness to relevant confounding factors.
Exploratory analyses may identify features warranting further investigation. An association observed in an exploratory dataset does not establish that a feature is a validated biomarker.
Important: Candidate-feature identification is an exploratory research step. Statistical association or predictive performance in an exploratory dataset does not establish clinical validity, clinical utility, or biomarker status. Independent validation is required before making such conclusions.
Features, thresholds, preprocessing decisions, and model specifications selected using exploratory or training data should not be evaluated on the same observations used to make those selections.
Where predictive modelling is performed, participant-level separation should generally be used so that recordings from the same participant do not appear in both development and validation datasets. This reduces the risk that measured performance reflects participant-specific memorization or within-participant leakage rather than generalization to new individuals.
This methodology describes the processing pipeline associated with Cogitrac Voice version 1.0.0. Changes to preprocessing parameters, feature definitions, recording eligibility criteria, reference-window rules, or statistical thresholds may produce results that are not directly comparable across versions.
The production implementation and its version-controlled pipeline specifications are the definitive references for exact processing behavior.
For reproducible research, the applicable implementation record should identify, where relevant:
Version control: The placeholder dependency versions above should be replaced with the exact versions used in the production or research environment. If dependency versions are maintained in a lockfile, container, requirements manifest, or software release record, that artifact should be treated as part of the reproducibility record.
A comprehensive technical document with complete processing specifications is available upon request. Subject to applicable confidentiality and data-governance requirements, it may include:
Complete technical details are available to qualified researchers and collaborators where appropriate.
Contact Research TeamReference scope: These publications provide background on acoustic voice measurement, software, and signal processing. They do not constitute validation of the Cogitrac Voice feature set for epilepsy, seizures, or other neurological outcomes.