In total, 19 English-speaking, cis-gender female participants were recruited from Klick Health, a technology, media, and research company in the health care sector based in Toronto, Canada. Recruitment information was posted on the company’s internal communication platform, and interested parties contacted our research team privately for more information and to sign the informed consent. After the informed consent was signed, participants were onboarded through either a video call or in person. As part of the onboarding, participants were provided instructions on how to conduct the urine tests, take their BBT, and record their voice through the research app.

All participants reported a natural menstrual cycle without taking hormonal birth control or medication that would affect the menstrual cycle and reported their menstrual cycle as consistent (1‐2 days variation per cycle) or somewhat consistent (3‐4 days variation per cycle). Participants were instructed to take their BBT, self-administer a consumer-grade LH urine test (Easy@Home Ovulation Tests; Easy@Home Fertility), and record themselves saying “Hello, how are you?” every morning immediately upon waking for 1 full menstrual cycle. The decision to use a fixed phrase was made based on the success of previous studies in using fixed sentences and conversational speech to identify significant features between menstrual phases compared to vowel sounds [3,8,15]. Participants were instructed to record their voices in a quiet environment with no other voices present in the recording and took pictures of the LH test strips each day to record hormone test results.

The data collection started the day after the end of menstruation, and recordings were collected for 1 full menstrual cycle, ending on the last day of menstruation. Participants did not perform the hormone test on days they were menstruating but did record their voice and BBT. Data (LH test results, BBT, and voice recordings) were collected using a custom mobile app developed by Klick Applied Science that participants installed on their personal cellphones following the procedure below:

Participants were excluded if the LH peak could not be detected, either through multiple consecutive missed data recordings (n=2) or if no discernable LH rise was detected through hormone tests (n=1). In total, 16 participants were included with a total of 320 voice recordings.

Ethics was obtained from the Canadian SHIELD Ethics Review Board (REB tracking 2022-02-004). Informed consent was collected for each participant, and all procedures corresponded to local guidelines. Participant data were deidentified during collection. Participants were compensated US $77 for participating in the study and received an additional US $38.50 if they recorded all the voice recordings.

Individual participants recorded their voices once a day, every day for 1 menstrual cycle. The average menstrual cycle length was 28.81 (SD 2.90) days. The average length of the follicular phase (including menstruation and the day of the LH surge and the day after) was 15.92 (SD 3.30) days, and the average length of the luteal phase was 12.75 (SD 1.44) days. On average, menstruation was 5.54 (SD 1.66) days and ended 9.44 (SD 3.12) days prior to the LH surge. In total, 320 voice recordings were collected. Study participants had an average age of 30.6 (SD 4.5) years and an average BMI of 22.7 (SD 2.8) kg/m² and included various ethnic backgrounds (East or Southeast Asian: n=4, 25%; Jewish: n=1, 6%; Middle Eastern: n=2, 13%; South Asian: n=6, 37%; White: n=3, 19%; self-identification).

The LH surge was defined to be when the hormone strip test line was at its maximum darkness out of all the days in the menstrual cycle and approximately equal to or greater than the darkness of the control line. The ovulation window was defined as the day of the LH peak and the day after (to cover the 24‐36 hour likelihood for ovulation). Days prior to this were classified as “follicular,” corresponding to the follicular phase in the menstrual cycle, and days following ovulation classification and occurring prior to menstruation were classified as “luteal.” Recordings collected while participants were menstruating were labeled “menstrual.” BBT was used to verify the likelihood that ovulation occurred through a sustained 0.3 °C increase in the days following the suspected ovulation day.

Each audio recording was assessed by researchers to ensure adequate recording quality and the absence of background noise. In total, 4 vocal features associated with the recording F0 were extracted from each voice recording: the mean (F0AV), SD (F0SD), 5th percentile (F0MIN), and 95th percentile (F0MAX). The 5th and 95th percentiles were chosen as robust analogs to the minimum and maximum F0 values. Voice features were extracted using librosa [17], a Python library widely used for music and audio analysis, providing tools for audio feature extraction and processing tasks. F0 values for the recording were extracted using the pyin function, with the specified frequency minimum of C2 (~65 Hz) and maximum of C7 (~2093 Hz). These values were chosen to encompass a wide range of vocal frequencies and are the recommended minimum and maximum frequencies for analysis [17].

Each voice feature was scaled independently for each participant. Each voice feature was minimum-maximum scaled according to the following formula:

Where X_sc is the scaled feature value, X is the original feature value, X_min is the minimum daily value of the feature throughout the participant’s menstrual cycle, and X_max is the maximum daily value of the feature throughout the participant’s menstrual cycle. This scaling was conducted such that the maximum scaled feature value throughout the recording period was 1 and the minimum value was 0 (Figure 1). This was repeated for each feature, for each participant.

To assess the change in voice features throughout the menstrual cycle, we used changepoint detection. Changepoint detection algorithms are tools used to identify points in time where the statistical properties of a data sequence change. These changes can signify shifts in mean, variance, distribution, or other characteristics of the data and can be used to detect shifts in patient health status or the effectiveness of treatments over time [18-23]. Changepoint detection algorithms are especially valuable for identifying transitions between distinct stages [23], making them a promising tool for research on physiological changes throughout the menstrual cycle. In this study, we used the Ruptures package, a Python library designed for offline changepoint detection, which considers all data points within a time series to identify underlying statistical changes in the data [24]. This unsupervised algorithm does not use prior knowledge of ovulation days and is not influenced by other participants’ data, ensuring that the detection process only identifies changes in the voice recording F0 data themselves.

We applied 3 changepoint detection algorithms: dynamic programming, binary segmentation, and bottom-up segmentation. Dynamic programming explores all possible partitions of the time series to identify the optimal segmentation, constrained to detect only a single changepoint. Binary segmentation iteratively identifies the most significant changepoint in each segment but, for this study, was configured to stop after detecting the first changepoint. Bottom-up segmentation begins with many small segments and merges the segments that are most similar until only 1 changepoint remains, meeting the defined stopping criterion.

To quantify changes, we used 3 cost functions: least absolute deviation loss function (L1), least squared deviation loss function (L2), and radial basis function. The least absolute deviation cost function measures absolute differences, making it robust to outliers. The least squared deviation cost function emphasizes larger deviations by using squared differences. The radial basis function cost function captures nonlinear changes by comparing segments with a radial basis kernel. Detailed descriptions of these algorithms and cost functions are available in the Ruptures documentation [24].

We were not aiming to identify changes associated with menstruation so we specified the detection of a single changepoint within a window spanning from 9 days before to 9 days after the LH surge. These time points correspond to the average number of days we observed occurring prior to the LH surge following menstruation (9.4, SD 3.1 days) and the postovulatory day that the corpus luteum begins to disintegrate and hormone levels fall [25]. The algorithm was configured with a minimum distance of 3 days between any 2 breakpoints or from the start of the time series, with a jump parameter set to 1 to ensure that each day was evaluated for potential changepoints of the vocal features between menstrual phases.

To analyze differences between the follicular and luteal phases, 4 generalized linear mixed models were applied for the 4 F0 features. The voice feature was the dependent variable, the menstrual phase (follicular vs luteal) was the independent variable, and participant ID was a random effect to account for intraindividual variability. To avoid data that may occur between the follicular and luteal phases and ensure the clarity of phase-specific effects, days identified as ovulatory (the day of LH surge and the following day) and menstrual days were excluded from the analysis. All other data points were used. The models were estimated using restricted maximum likelihood.

To assess the significance of our changepoint findings against a null hypothesis of random results, we conducted 1-sided 2-proportion z tests. We compared the observed proportion of participants with a changepoint occurring before ovulation in the fertile window (days −5 to +1 relative to LH surge) to a random expectation, corresponding to 50% (n=8) of participants with a changepoint occurring in the fertile window. This proportion is based on 7 days in the fertile window of an average of 14 days in the possible prediction window (6 days prior to the LH surge maximum, the LH surge maximum, and 7 days following the LH surge maximum). The alternate hypothesis posited that the observed proportion would be higher than the random expectation due to the hormonal changes associated with the fertile window. All statistical analysis was performed in Python (version 3.12.2; Python Software Foundation) using the statsmodels package (mixedlm and proportions_ztest functions) [26].

To assess vocal dynamics consistent across participants, generalized linear mixed models were fit for each of the F0 features (F0AV, F0SD, F0MIN, and F0MAX). Overall, there was a significant difference in the F0SD and F0MIN between the follicular and luteal recording days for each participant, such that F0SD was, on average, 9.0% (SD 2.9%) lower in the luteal phase compared to the follicular phase (95% CI 3.4%‐14.7%; P=.002) and F0MIN was 8.8% (SD 3.6%) higher (95% CI 1.7%‐16.0%; P=.01). There were no significant differences between phases in F0AV or F0MAX (P=.65 and P=.07). Average daily trends of the participant-scaled F0 values are displayed in Figure 2.

F0 feature data for each participant was fed into the changepoint detection algorithms, and the predicted changepoints are displayed in Figure 3. Menstrual phases were not added as information in the algorithms at any point, and any changepoints detected solely reflect changes in the vocal features themselves. Additionally, data from each participant were used independently for changepoint identification, ensuring that data were not shared or mixed between participants. Overall, the average (SD) detected changepoints for all algorithms were 0.20 (4.13) days following ovulation in F0AV, 1.42 (3.52) days prior to ovulation in F0SD, 0.88 (3.27) days prior to ovulation in F0MIN, and 1.31 (4.05) days following ovulation in F0MAX. As the variation in the predicted changepoints was large, we also looked where the predicted changepoints occurred specifically. We observed that up to 81% (n=13) of participants had a predicted changepoint occurring within the fertile window, or the 5 days preceding the LH surge to the 1 day following the LH surge, for both F0SD and F0MIN (P=.03). Meanwhile, a maximum of 63% (n=10; P=.24) and 50% (n=8; P=.50) of participants had changepoints detected in the fertile window for F0AV and F0MAX, respectively (Figure 3).

This study explored how vocal pitch features fluctuate across the menstrual cycle, with a focus on vocal patterns surrounding phase transitions and ovulation. Building on prior research [15], we expanded the analysis by examining a broader set of F0-related features, including minimum and maximum pitch values, using real-world voice recordings collected daily via smartphones. With more frequent recordings, we investigated whether these vocal features varied between the follicular and luteal phases, providing a more comprehensive perspective on daily vocal behavior. We further hypothesized that changepoint detection could pinpoint the transition between these phases, offering insight into the potential for voice-based markers of fertility in future applications.

Physiologically, the differences in voice features between menstrual phases may be attributed to differences in hormone concentrations. In the follicular phase, estrogen levels rise, whereas the luteal phase is characterized by increased progesterone levels. The larynx itself is a sexual organ and can be influenced by reproductive hormones, as estrogen stimulates the mucosal cells. Mucus on the vocal cords has been shown to be similar in consistency and makeup as cervical mucus throughout the menstrual cycle, displaying increased thin, wet mucus in the follicular phase and thick, dry mucus in the luteal phase [27]. A thin, wet mucus layer is necessary for proper vocal cord vibration and range of pitch [28], and decreased vocal range has been associated with thicker, dry laryngeal mucus [27,29]. Our results align well with these depictions, as we observed a decrease in the F0SD and range (observed through an increase in the minimum F0) in the luteal phase. An alternative explanation for vocal changes may reside in more nuanced apps. Fatigue and stress may be exacerbated in the luteal phase [30], and increased fatigue has been associated with increased F0SD [31]. Regardless, changes in mental and hormonal states associated with the changing menstrual phases provide evidence for the observed vocal changes.

We found that individually scaled F0 voice features had distinct differences between the follicular and luteal phases. Similar to what was observed in previous studies, F0SD decreased and F0MIN increased in the luteal phase [14,15]. Meanwhile, the association between the F0AV and the menstrual phase was insignificant. F0AV throughout the menstrual cycle has traditionally reported mixed results, in which some authors found that F0 increased in periods of high fertility [8,32], decreased in high fertility [15,33,34], or did not change [6,9,14,35-39]. This difference may be attributed to different study methodologies, such as recording a vowel instead of a fixed phrase or only recording 1 voice sample per menstrual phase.

Changes in the voice surrounding the fertile period have been hypothesized to be the result of evolutionary processes. More attractive female participants have an increased likelihood of finding a suitable mate, and mating during the fertile window has the highest likelihood of conception [1]. Research has shown that variations in female voice characteristics throughout the menstrual cycle can influence perceived attractiveness. In the high fertility days prior to ovulation, female voices have typically been rated “more attractive” by heterosexual male participants [15,16]. It has been shown that 44%‐55% of the variance in the desirability scores for female participants was explained by their minimum F0 in speech. This may indicate that female voices with lower pitch minima, opposed to a change in the mean pitch, are more desirable [40]. Our findings align nicely with this result, as the 5th percentile of the pitch was an average of 8.8% lower in the follicular phase (a phase corresponding to higher fertility) compared to the luteal phase for all our participants. Furthermore, increased pitch variability, similar to what was seen in the F0SD in the follicular phase, has been associated with increased social attractiveness of the speaker [41,42].

One of the primary distinctions of our work is the employment of a changepoint detection algorithm to identify the day of the switch in vocal features within the menstrual cycle. Although the average predicted changepoints fell within 2 days of ovulation in all participants, there was a large variation in the values that would not permit accurate ovulation prediction. However, when assessing the changepoints in the broader sense of fertility identification, we note that changepoints are identified within the fertile window in a high proportion of participants. The fertile window is associated with an increased likelihood of conception, making it a meaningful target for detection, even if exact ovulation timing is not achieved [1]. In particular, we observed that changepoints are detected within the fertile window in up to 81% (n=13) of participants for F0SD and F0MIN. However, this result was not observed for F0AV and F0MAX.

Offline changepoint detection is a useful method for fertility monitoring, even without the real-time benefits of online detection (in which changepoints can be detected as the data are collected). Current fertility monitoring methods can also require daily use or cycle-long data collection to accurately pinpoint the timing of ovulation. For example, BBT is used as an ovulation confirmation, in which the BBT must be measured every day for at least 1 full menstrual cycle to identify ovulation [1]. After further refinement, offline detection used in a consumer product would have similar confirmatory utility.

Unlike previous studies that had limited voice data collection methodologies, our approach reflects real-world situations by using at-home smartphone recordings, allowing for a more comprehensive assessment of voice parameters throughout the menstrual cycle. This opens up avenues for exploring the feasibility of using voice parameters as indicators of high fertility and ovulation. Moreover, the simplicity and noninvasiveness of smartphone-based recording make it suitable for widespread use in both clinical settings and personal health monitoring. With just a voice and a smartphone app, users do not require any additional devices, highlighting the convenience and accessibility of this method.

Our study’s insights into vocal pitch changes across the menstrual cycle are limited by a small sample size (n=16 participants), which may not capture the full spectrum of variability across a wider population. Additionally, analyzing recordings from only 1 menstrual cycle per participant neglects potential intraindividual variability across different cycles. Finally, our examination of a limited set of pitch features restricts our understanding of the nuanced relationship between vocal characteristics and menstrual phases, and there may be more information located in other pitch-related features. With further refinement through larger sample sizes, multicycle analyses, and a broader examination of vocal features, we could gain better insights into how vocal pitch correlates with menstrual cycle changes. Importantly, using larger, multicycle datasets could enable the use of online changepoint detection algorithms, which have the advantage of identifying phase transitions in real-time without requiring access to the entire dataset beforehand, or could be used to train a machine learning model for enhanced performance. With these advancements, this method could contribute to the early detection of fertility issues, personalized fertility tracking, and natural family planning.

Our study contributes to the evolving understanding of how menstrual cycle phases impact vocal pitch, underscoring the potential of voice as a noninvasive marker for fertility and ovulation. Using methods that use real-world data captured through smartphone recordings, our study analyzed these data with changepoint detection algorithms to identify transitions between menstrual phases. This approach not only confirms the relationship between vocal pitch variations and the menstrual cycle phase but also demonstrates the feasibility of applying such findings in practical, real-world settings. Our work could have implications for clinical and personal health monitoring, offering potential improvements in fertility tracking and natural family planning methods. Future efforts to integrate more sophisticated machine learning techniques and real-time analysis could further refine the accuracy and applicability of vocal pitch as a tool for women’s health management.