Elina Helander

Voice Conversion Using Partial Least Squares Regression

Voice conversion can be formulated as finding a mapping function which transforms the features of the source speaker to those of the target speaker. Gaussian mixture model (GMM)-based conversion is commonly used, but it is subject to overfitting. In this paper, we propose to use partial least squares (PLS)-based transforms in voice conversion. To prevent overfitting, the degrees of freedom in the mapping can be controlled by choosing a suitable number of components. We propose a technique to combine PLS with GMMs, enabling the use of multiple local linear mappings. To further improve the perceptual quality of the mapping where rapid transitions between GMM components produce audible artefacts, we propose to low-pass filter the component posterior probabilities. The conducted experiments show that the proposed technique results in better subjective and objective quality than the baseline joint density GMM approach. In speech quality conversion preference tests, the proposed method achieved 67% preference score against the smoothed joint density GMM method and 84% preference score against the unsmoothed joint density GMM method. In objective tests the proposed method produced a lower Mel-cepstral distortion than the reference methods.

Voice Conversion Using Dynamic Kernel Partial Least Squares Regression

A drawback of many voice conversion algorithms is that they rely on linear models and/or require a lot of tuning. In addition, many of them ignore the inherent time-dependency between speech features. To address these issues, we propose to use dynamic kernel partial least squares (DKPLS) technique to model nonlinearities as well as to capture the dynamics in the data. The method is based on a kernel transformation of the source features to allow non-linear modeling and concatenation of previous and next frames to model the dynamics. Partial least squares regression is used to find a conversion function that does not overfit to the data. The resulting DKPLS algorithm is a simple and efficient algorithm and does not require massive tuning. Existing statistical methods proposed for voice conversion are able to produce good similarity between the original and the converted target voices but the quality is usually degraded. The experiments conducted on a variety of conversion pairs show that DKPLS, being a statistical method, enables successful identity conversion while achieving a major improvement in the quality scores compared to the state-of-the-art Gaussian mixture-based model. In addition to enabling better spectral feature transformation, quality is further improved when aperiodicity and binary voicing values are converted using DKPLS with auxiliary information from spectral features.

Factors related to sustained use of a free mobile app for dietary self-monitoring with photography and peer feedback: retrospective cohort study.

Background Healthy eating interventions that use behavior change techniques such as self-monitoring and feedback have been associated with stronger effects. Mobile apps can make dietary self-monitoring easy with photography and potentially reach huge populations. Objective The aim of the study was to assess the factors related to sustained use of a free mobile app (“The Eatery”) that promotes healthy eating through photographic dietary self-monitoring and peer feedback. Methods A retrospective analysis was conducted on the sample of 189,770 people who had downloaded the app and used it at least once between October 2011 and April 2012. Adherence was defined based on frequency and duration of self-monitoring. People who had taken more than one picture were classified as “Users” and people with one or no pictures as “Dropouts”. Users who had taken at least 10 pictures and used the app for at least one week were classified as “Actives”, Users with 2-9 pictures as “Semi-actives”, and Dropouts with one picture as “Non-actives”. The associations between adherence, registration time, dietary preferences, and peer feedback were examined. Changes in healthiness ratings over time were analyzed among Actives. Results Overall adherence was low—only 2.58% (4895/189,770) used the app actively. The day of week and time of day the app was initially used was associated with adherence, where 20.28% (5237/25,820) of Users had started using the app during the daytime on weekdays, in comparison to 15.34% (24,718/161,113) of Dropouts. Users with strict diets were more likely to be Active (14.31%, 900/6291) than those who had not defined any diet (3.99%, 742/18,590), said they ate everything (9.47%, 3040/32,090), or reported some other diet (11.85%, 213/1798) (χ2 3=826.6, P<.001). The average healthiness rating from peers for the first picture was higher for Active users (0.55) than for Semi-actives (0.52) or Non-actives (0.49) (F 2,58167=225.9, P<.001). Actives wrote more often a textual description for the first picture than Semi-actives or Non-actives (χ2 2=3515.1, P<.001). Feedback beyond ratings was relatively infrequent: 3.83% (15,247/398,228) of pictures received comments and 15.39% (61,299/398,228) received “likes” from other users. Actives were more likely to have at least one comment or one “like” for their pictures than Semi-actives or Non-actives (χ2 2=343.6, P<.001, and χ2 2=909.6, P<.001, respectively). Only 9.89% (481/4863) of Active users had a positive trend in their average healthiness ratings. Conclusions Most people who tried out this free mobile app for dietary self-monitoring did not continue using it actively and those who did may already have been healthy eaters. Hence, the societal impact of such apps may remain small if they fail to reach those who would be most in need of dietary changes. Incorporating additional self-regulation techniques such as goal-setting and intention formation into the app could potentially increase user engagement and promote sustained use.

A Novel Method for Prosody Prediction in Voice Conversion

Most of the published voice conversion schemes do not consider detailed prosody modeling but only control the F0 level and range. However, the detailed prosody can also carry a significant amount of speaker identity related information. This paper introduces a new method for converting the prosody in voice conversion. A syllable-based prosodic codebook is used to predict the converted F0 using not only the source contour but also linguistic information and segmental durations. The selection of the most suitable target contour is carried out using a trained classification and regression tree. The F0 contours in the codebook are represented in a transformed domain which allows compression and fast comparison. The performance of the prosodic conversion is evaluated in a real voice conversion system. The results indicate a significant improvement in speaker identity and naturalness when compared to GMM (Gaussian mixture model) based pitch prediction approach.

LSF mapping for voice conversion with very small training sets

To make voice conversion usable in practical applications, the number of training sentences should be minimized. With traditional Gaussian mixture model (GMM) based techniques small training sets lead to over-fitting and estimation problems. We propose a new approach for mapping line spectral frequencies (LSFs) representing the vocal tract. The idea is based on inherent intra-frame correlations of LSFs. For each target LSF, a separate GMM is used and only the source and target LSF elements best correlating with the current LSF are used in training. The proposed method is evaluated both objectively and in listening tests, and it is shown that the method outperforms the conventional GMM approach especially with very small training sets.

Are Breaks in Daily Self-Weighing Associated with Weight Gain?

Regular self-weighing is linked to successful weight loss and maintenance. However, an individuals self-weighing frequency typically varies over time. This study examined temporal associations between time differences of consecutive weight measurements and the corresponding weight changes by analysing longitudinal self-weighing data, including 2,838 weight observations from 40 individuals attending a health-promoting programme. The relationship between temporal weighing frequency and corresponding weight change was studied primarily using a linear mixed effects model. Weight change between consecutive weight measurements was associated with the corresponding time difference (β = 0.021% per day, p<0.001). Weight loss took place during periods of daily self-weighing, whereas breaks longer than one month posed a risk of weight gain. The findings emphasize that missing data in weight management studies with a weight-monitoring component may be associated with non-adherence to the weight loss programme and an early sign of weight gain.

Objectively measured physical activity in Finnish employees: a cross-sectional study

Objectives To objectively measure the amount of intensity-specific physical activity by gender and age with respect to body mass index (BMI) during workdays and days off among Finnish employees. Design A cross-sectional study. Setting Primary care occupational healthcare units. Participants A sample of 9554 Finnish employees (4221 men and 5333 women; age range 18–65 years; BMI range 18.5–40 kg/m2) who participated in health assessments related to occupational health promotion. Main outcome measurements The amount of moderate-to-vigorous (MVPA) and vigorous (VPA) physical activity (≥3 and ≥6 metabolic equivalents, respectively) was assessed by estimating the minute-to-minute oxygen consumption from the recorded beat-to-beat R-R interval data. The estimation method used heart rate, respiration rate and on/off response information from R-R interval data calibrated by age, gender, height, weight and self-reported physical activity class. The proportion of participants fulfilling the aerobic physical activity recommendation of ≥150 min/week was calculated on the basis of ≥10 min bouts, by multiplying the VPA minutes by 2. Results Both MVPA and VPA were higher among men and during days off, and decreased with increasing age and BMI (p<0.001 for all). Similar results were observed when the probability of having a bout of MVPA or VPA lasting continuously for ≥10 min per measurement day was studied. The total amount of VPA was low among overweight (mean ≤2.6 min/day), obese (mean ≤0.6 min/day) and all women in the age group 51–65 years (mean ≤2.5 min/day) during both types of days. The proportion of participants fulfilling the aerobic physical activity recommendation was highest for normal weight men (65%; 95% CI 62% to 67%) and lowest for obese women (10%; 95% CI 8% to 12%). Conclusions Objectively measured physical activity is higher among men and during days off, and decreases with increasing age and BMI. The amount of VPA is very low among obese, overweight and older women.

Physical Activity: Absolute Intensity versus Relative-to-fitness-level Volumes

Purpose This study aimed to investigate in a real-life setting how moderate- and vigorous-intensity physical activity (PA) volumes differ according to absolute intensity recommendation and relative to individual fitness level by sex, age, and body mass index. Methods A total of 23,224 Finnish employees (10,201 men and 13,023 women; ages 18–65 yr; body mass index = 18.5–40.0 kg·m−2) participated in heart rate recording for 2+ d. We used heart rate and its variability, respiration rate, and on/off response information from R-R interval data calibrated by participant characteristics to objectively determine daily PA volume, as follows: daily minutes of absolute moderate (3–<6 METs) and vigorous (≥6 METs) PA and minutes relative to individual aerobic fitness for moderate (40%–<60% of oxygen uptake reserve) and vigorous (≥60%) PA. Results According to absolute intensity categorization, the volume of both moderate- and vigorous-intensity PA was higher in men compared with women (P < 0.001), in younger compared with older participants (P < 0.001), and in normal weight compared with overweight or obese participants (P < 0.001). When the volume of PA intensity was estimated relative to individual fitness level, the differences were much smaller. Mean daily minutes of absolute vigorous-intensity PA were higher than those of relative intensity minutes in normal weight men ages 18–40 yr (17.7, 95% confidence interval [CI] = 16.9–18.6, vs 8.6, 95% CI = 8.0–9.1; P < 0.001), but the reverse was the case for obese women ages 41–65 yr (0.3, 95% CI = 0.2–0.4, vs 7.8, 95% CI = 7.2–8.4; P < 0.001). Conclusion Compared with low-fit persons, high-fit persons more frequently reach an absolute target PA intensity, but reaching the target is more similar for relative intensity.

Joint utilization of incoherently and coherently integrated radar signal in helicopter categorization

Radar signal based helicopter categorization is a challenging task for all types of radars. Airborne pulse Doppler radar with an appropriate digital signal processing unit has a good potential to perform categorization or even classification, providing that radar parameters are carefully selected. This paper presents a helicopter categorization method, which is based on estimation of the main rotor blade tip velocity and the time interval between two successive main rotor blade flashes. Both incoherent integration and conventional coherent integration play an important role in the new method. Joint utilization of both integration types makes the method feasible without using unrealistic radar waveform parameters. Based on the L/N-quotient helicopter classification method, a new categorization procedure was developed. Compared to the original L/N-quotient method, the new method exploits both incoherent and coherent integration. Combined use of both integration types enables it to use effectively potential of dwell time and PRF. In addition, the new method includes a certain kind of moving windows edge detection algorithm that was found to improve the performance of the edge detection compared to the one introduced with the original L/N-quotient method.

Evaluation of detailed modeling of the LP residual in statistical speech synthesis

Speech parameterization remains an open question in statistical speech synthesis. In our earlier work we have shown that a framework developed originally for highly efficient speech storage can also be successfully applied for voice conversion and concatenative unit selection based speech synthesis. Recently, we have also used the same coding scheme in hybrid-form speech synthesis. In this paper, we further discuss the framework and apply it in statistical speech synthesis, concentrating specifically on the spectral modeling of the linear prediction (LP) residual. Perceptual evaluation demonstrates that the modeling of the spectral details remaining in the residual improves the quality of synthetic speech.