It's not clear what that figure is based on, but there is one paper which gives the "70%" number in its abstract, but as:
The use of efficacious face-masks (such as surgical masks, with estimated efficacy >= 70%) in public could lead to the elimination of the pandemic if at least 70% of the residents of New York state use such masks in public consistently (nationwide, a compliance of at least 80% will be required using such masks). The use of low efficacy masks, such as cloth masks (of estimated efficacy less than 30%), could also lead to significant reduction of COVID-19 burden (albeit, they are not able to lead to elimination)
And from the body of the paper, some supporting data, but note that it's not Covid-19 specific:
We estimated the efficacy of face-masks based on the results of a number of clinical trials. For instance, data from Driessche et al.  shows that surgical masks reduced P. aeruginosa infected aerosols produced by coughing by over 80% in cystic fibrosis patients. A similar study by Stockwell et al.  shows that surgical masks reduced colony-forming unit (CFU) count by over 90% (these two studies in ,  show that the N95 masks (respirators) were more effective). Similarly, van der Sande et al.  show that home-made tea-cloth masks had an inward efficiency between 58% and 77% over a 3-hour duration of wear, while inward efficiency ranged 72%–85% and 98%–99% for surgical and N95-equivalent masks. Consequently, following Eikenberry et al. , we estimate inward mask efficacy to range widely between 20%–80% for cloth masks, and at least 50% for well-made, tightly fitting masks made of optimal materials, and 70%–90% for surgical masks, and over 95% typical for p in the range of properly worn N95 masks.
There are some more recent papers on mask efficiency in common respiratory viruses, but e.g. Cowling et al. (2020) the one below presents its data as P-values for a Tobit regression. I don't know how to convert this to an efficacy % number. (Also "coronavirus" refer to common-cold coronaviruses in the quote.)
a–c, Virus copies per sample collected in nasal swab (red), throat swab (blue) and respiratory droplets collected for 30min while not wearing (dark green) or wearing (light green) a surgical face mask, and aerosols collected for 30min while not wearing (brown) or wearing (orange) a face mask, collected from individuals with acute respiratory symptoms who were positive for coronavirus (a), influenza virus (b) and rhinovirus (c), as determined by RT–PCR in any samples. P values for mask intervention as predictor of log10 virus copies per sample in an unadjusted univariate Tobit regression model which allowed for censoring at the lower limit of detection of the RT–PCR assay are shown, with significant differences in bold. For nasal swabs and throat swabs, all infected individuals were included (coronavirus, n=17; influenza virus, n=43; rhinovirus, n=54). For respiratory droplets and aerosols, numbers of infected individuals who provided exhaled breath samples while not wearing or wearing a surgical face mask, respectively were: coronavirus (n=10 and 11), influenza virus (n=23 and 28) and rhinovirus (n=36 and 32). A subset of participants provided exhaled breath samples for both mask interventions (coronavirus, n=4; influenza virus, n=8; rhinovirus, n=14). The box plots indicate the median with the interquartile range (lower and upper hinge) and ±1.5×interquartile range from the first and third quartile (lower and upper whiskers).
I should add that there's a more recent (alas just theoretical study that highlights the duality between mask efficiency and adherence to mask use, in relation to the reduction in R0 (this is more relevant to the highlighted claim from the "duplicate" question "If everybody is wearing masks you eliminate the risk of spread of Covid-19 by 98.5 per cent":
A conservative assessment applied to the COVID-19 estimated R0 of 2.4 (7) might posit 50% mask usage and a 50%
mask efficacy level, reducing R0 to 1.35, an order of magnitude
impact rendering spread comparable to the reproduction number of seasonal influenza. To put this in perspective, 100 cases
at the start of a month becomes 31,280 cases by the month’s
end (R0 = 2.4) vs. only 584 cases (R0 = 1.35). Such a slowdown in case-load protects healthcare capacity and renders a
local epidemic amenable to contact tracing interventions that
can eliminate the spread entirely.
A full range of efficacy e and adherence pm is shown with
the resulting R0 in Figure 1, illustrating regimes in which
growth is halted entirely (R0 < 1) as well as pessimistic
regimes (e.g. due to poor implementation or population compliance) that nonetheless result in a beneficial effect in suppressing the exponential growth of the pandemic.
But this study/review does mention a similar one (besides some cross-country studies, which I'm omitting here; for those see a related q):
Yan et al (90) provide an additional example of an incremental impact assessment of respiratory protective devices
using an augmented variant of a traditional SIR model in the
context of influenza with N95 respirators. They showed that
a sufficiently high adherence rate (~ 80% of the population)
resulted in the elimination of the outbreak with most respiratory protective devices.
- J Yan, S Guha, P Hariharan, M Myers, Modeling the Effectiveness of Respiratory Protective
Devices in Reducing Influenza Outbreak. Risk Analysis 39, 647–661 (2019).
A relevant graph from this latter (Yan et al.) study, which is also theoretical as far as the epidemiology goes, but uses actual filtration figures for the various kinds of respirators/masks, and actual breathing rate volumes (per person per day), so a bit more concrete in that sense. Also, Yan et al. don't assume an inherent R0, but use pathogen size as input for their modified SIR model, in that study the influenza virus size. (The influenza A virus is actually somewhat smaller than the coronaviruses. The rhinovirus is even smaller still. You can see from the Cowling et al. paper that this virus-size difference does seem to make a difference when it comes to passing though masks, experimentally.) On top of that
The probability p of infection by an inhaled pathogen was taken to be 0.052 (Li, Eisenberg, Spicknall, & Koopman, 2009; Stilianakis &
Drossinos, 2010) for both adults and children.
Fig. 1. Infection prevalence for (a) adult fit-tested respirators, (b) unfitted adult respirators, (c) high-filtration surgical masks, and
(d) low-filtration surgical masks. Compliance rates are: 0% (curve with highest peak), 20% (second highest peak), 50% (third highest
peak), and 80% (lowest peak).