The theory of learning Armenian is one we will have to build ourselves

Methods for teaching Armenian as a foreign language exist, but there is almost no research base under them: almost every work found on Eastern Armenian is a description of teaching experience with no control groups and no learner data (the exception is a 2013 dissertation with pronunciation figures, again without a comparison group), and the reform of Western Armenian teaching in the diaspora is described in programme documents and has not been evaluated by outcomes. Not a single study comparing ways of learning Armenian on learners was found: no experiments with a comparison group, no error corpus, no estimate of time to a given level. At the same time, the general science of second language acquisition is mature and offers stable footholds: comprehensible input is necessary but not sufficient; explicit explanation and practice help; spaced practice and retrieval practice are the most reliable techniques; functional morphology is the bottleneck for everyone. The native language matters, and a lot: in the largest sample (about 50,000 adults, 62 native languages) it explains 9–22% of the total variance in the speaking score. For a Russian speaker this means a specific map: case as an idea, the absence of gender and free word order come almost for free; aspirated consonants, the definite article, the mobile auxiliary and an opaque core vocabulary do not. Almost all of this map is derived from comparing structures and from analogue languages, not measured on learners. Hence the main conclusion: there is no ready-made theory, but there is material to assemble one from, and almost every one of its propositions can be stated as a testable hypothesis; they are given at the end.

How to read the tags. Every claim here has a level of verification. A claim with a link and no tag is taken from a primary source that was opened: from the full text or from the abstract. The tag [unverified] means that the source is known only from metadata, a search snippet or a secondary retelling. The tag [inference] means a piece of reasoning: from a structural comparison of Russian and Armenian, from data on other languages, or from general theory; it has not been tested on learners of Armenian. All Armenian examples in the text are taken from sources; there are no home-made examples here.

Methods exist, measurements do not: a field of practitioners without a shared literature

Eastern Armenian: teachers' experience instead of data

There is no established research field of "Armenian as a foreign language". There is a small and scattered practitioner literature, almost all of it from Armenian universities and almost all of it from 2015–2026. The authors themselves admit this: a 2016 article from the Armenian Pedagogical University reports that Soviet Armenian pedagogy treated Armenian only as a native language and that the problem "is not fully studied" (Muradyan 2016); a 2024 article on teaching American diplomats concludes that "there is no precise categorization for the Armenian language from the standpoint of national standards" (Semirjyan-Bekmezyan and Vardanyan 2024); a 2026 article calls Eastern Armenian "critically under-resourced in the pedagogical domain" despite a good linguistic description, and says that materials break off after the B1 threshold (Kagirova 2026, only the abstract was read). A 2019 article built on a survey of 45 Armenian linguists says the same from the inside: Armenian is widely taught as a foreign language, but there is no research and no methodological guide on how to teach it as a second language, and there are no Armenian proficiency levels (Sargsyan 2019).

The three articles whose reference lists were read in full cite neither one another nor earlier work on Armenian as a second language. Their theoretical anchors are the Soviet communicative method (Passov 1989), audiolingual pattern practice (Lado and Fries 1958), Shcherba and native-language teaching methodology; modern second language acquisition research is represented by two general textbooks in one of the articles. The word "effectiveness" stands in titles and abstracts as an assertion from experience: the 2026 RAU article with this word in its title, read in full, contains not a single measurement: the method is called effective because it is more accessible and more engaging (Rostomyan 2026). This is the mark of a community of practice without a cumulative literature, not of a research field [inference].

There are two modern works specifically about Russian-speaking learners, both from the same university. The first is a report on the experience of the Russian-Armenian University (RAU), where Armenian has been taught to Russian-speaking students for more than twenty years in a three-semester course (Iskandaryan 2025). The method is called contrastive, relying on Russian and on Shcherba, with communicative work added; the textbook is "Krunk Hayastani" by A. S. Markosyan (1998), which begins with an introductory phonetics course. The article contains no counts, no corpus and no test results; the difficulties are listed "from many years of experience". This report is valuable as the only source from the learner side, and below it is used exactly that way: as a teacher's testimony, not as a measurement.

The second is Zh. Rostomyan's 2013 dissertation on an intensive Armenian course built on knowledge of Russian; its summary was read, not the dissertation itself (Rostomyan 2013). The course was tried in 1998–2005 on pupils in grades 2–8 of the Russian-stream classes of twelve Yerevan schools and in 2005–2011 on foreign students at RAU. The order is: one and a half to two months of oral work without letters, by the end of which 900–1,300 words are known, then literacy, then reading; grammar runs alongside, without memorising terms or rules, and practice makes up 75–88% of the material. The author describes the main phonetic difficulty as a sound that is not heard: փ is perceived as п, ձ as ц, հ and ղ as х. A sound the learner produces in isolation falls apart in a sentence, and the schwa ը is still not mastered by the end of the course. There are figures: in the literacy stage the letters are fully learned, word spelling about half, word pronunciation 40–60%; correct pronunciation of the hard sounds in 100 students rose from about 75–80% to 90–98% over a year (the table is only partly legible in the scan). There is no comparison group, so this measures the course rather than testing the method. The author's textbook for students from CIS countries came out in 2009 [unverified].

There are few textbooks, and they are aimed at beginners. The only textbook whose preface was read is Dora Sakayan's course for English speakers (2007): a contrastive approach plus communicative competence, micro-dialogues with variable slots, grammar that "is not a goal in itself, but a means of achieving 'communicative competence'", romanization as an entry into the script (Sakayan 2007). Three independent sources (2014, 2024, 2026) agree that Eastern Armenian materials end at roughly A2–B1 and are overloaded with grammar. The Soviet line is visible through references: A. Gharibyan's "Short Course of the Armenian Language", built on comparison with Russian, came out six times from 1944, and from 1950 Gharibyan wrote grammar textbooks for the non-Armenian schools of Soviet Armenia (Margaryan 1989). The textbooks by Parnasyan and Manukyan and by Gharibyan, the Assimil course, and the methods of the textbooks by Bardakjian and Thomson, Samuelian and Hakobyan were not opened in this work: their existence is either confirmed only bibliographically or not confirmed at all [unverified]. The common denominator of all the traditions that could be seen is comparison with the learner's native language and an introductory course in script and phonetics as a separate stage; teaching Armenian monolingually as a matter of principle does not appear in the sources.

There was no state-level framework before 2024. The programme "Armenian for non-Armenian speakers: teaching, testing, assessment" was launched by an order of the Minister of Education dated 26 March 2024; it relies on the Common European scale and introduces six levels from Ա1 to Գ2 (Language Committee). In December 2024, standards for the four skills and thematic vocabulary minimums for A1–A2, B1 and B2 were presented; for C1–C2 only general characteristics are given, and the test and the platform were announced as pilots (Izmirlian Foundation). The texts of the standards and vocabulary minimums could not be found in open access, and as of October 2026 there is no evidence of a working exam either. The information that for citizenship the test on the Constitution in effect counts as the language test comes from a single commercial source with contradictory figures [unverified].

Western Armenian: the opposite stance and the same shortage of outcomes

Western Armenian is taught mainly in diaspora day schools as a second language to bilingual children. There is no monolingual community of its speakers at all: all speakers are a bilingual minority (Kelly and Keshishian 2021). Since 2015 the Armenian department of the Gulbenkian Foundation has been running a reform that explicitly rejects teaching built on grammar and language purity: the language should be "a real, used (and why not 'abused'), 'breathing' language and not a set of ossified rules", and should develop in a decentralized way (Gulbenkian, 2015 conference outcomes). Its practical embodiment is the Zarmanazan immersion programme, running since 2017: art, theatre, writing and games in the language, plus a university certificate for teachers jointly with INALCO (Gulbenkian, Zarmanazan 2023). No evaluation of Zarmanazan's outcomes was found: level before and after, retention, changes in teachers' practice.

The two traditions solve different problems with opposite default stances [inference]. The Western one: children with partial home input, the goal is use and identity, the method is immersion and production, grammar is deliberately pushed back. The Eastern one: literate adults with zero input, the goal is integration or study, the method is contrastive explanation through the native language with dialogues added. The Western side has something the Eastern side lacks: an explicitly stated theory of why the old method does not work. What both lack is the same thing: outcome data. Nobody has compared them with each other or with a control.

The little that is known empirically about non-native Armenian was obtained from heritage speakers and is concentrated in phonetics. In heritage speakers of Western Armenian, English shifts only those vowels that are already close to English ones, and the shift also occurs in those who learned English as adults (Godson 2004). In Lebanon speakers contrast voiced and voiceless unaspirated stops on the model of Lebanese Arabic; in the US the picture is variable, up to an English one (Kelly and Keshishian 2021). The only study of Eastern Armenian heritage speakers is a qualitative dissertation based on interviews in Los Angeles, which described a cycle of anxiety: shame over "broken" Armenian leads to avoidance, and avoidance leads to lower proficiency (Karapetian 2014). Whether a similar mechanism operates on adult foreigners in Yerevan has not been studied.

Digital tools and hours: supply without evidence

The AGBU online school (both varieties of the language) and the apps AYOlingo and Ariboon exist, but all the available figures about them are reach and volume of content, not learning outcomes, and they come from app stores and press releases [unverified]; no independent evaluation of any tool was found. One small point remained unresolved: on one reading, the RAU article names specific apps, including Duolingo; on another, it recommends apps without naming them, and the absence of an Armenian course on Duolingo is known only from secondary sources [unverified].

The only institutional estimate of time is American, for diplomats, and its figure exists in three versions. The current State Department page, opened on 13 October 2026 at a new address, places Armenian in Category III ("hard languages") and gives about 44 weeks, or 1,012 class hours, to level 3 on the ILR scale in speaking and listening, with 23 hours of classes and 17 hours of self-study per week (state.gov). An archived copy of the same page from 2024 gives 44 weeks and 1,100 hours for the same category, with the goal named as level 3 in speaking and reading (archive); the same figure of 1,100 is repeated by the article on teaching diplomats (Semirjyan-Bekmezyan and Vardanyan 2024). As of 13 October 2026 the old address of the page returns a 404 error, and in the table of the classic article on the institute's fifty years of experience Armenian is not named: the list of languages in the 44-week group breaks off with the word "etc." (Jackson and Kaplan 1999). The versions are compatible: 44 weeks at 23 hours give 1,012, and at 25 hours give 1,100, so the discrepancy looks like a change of weekly load in one and the same table [inference]. Something else matters more: all three figures were measured on high-aptitude native speakers of English studying full time, and Russian itself sits in the same category. They say nothing about the distance from Russian to Armenian; no estimate for Russian speakers exists.

Against this background, courses in Armenia itself are short: 48 hours at the American University of Armenia, about 45 hours at the state Repatriation and Integration Center, 4 hours a week at Birthright (Repat Armenia). That is 4–5% of the American estimate of the path to a professional level [inference], arithmetic on the numbers given. No descriptions of courses for Russian-speaking relocants after 2022 turned up in the sources.

The science of language acquisition: what has stood up to testing

Six disciplines and not one complete theory

Second language acquisition is studied not by one science but at the meeting place of several. Linguistics describes what exactly is acquired; cognitive psychology, the mechanisms of learning and memory; psycholinguistics, real-time processing; neuroscience, memory systems and plasticity; pedagogy, experiments with instruction; sociolinguistics, who gets access to input and why they should keep going. The standard collection of theories came out in its fourth edition in 2025 and contains chapters on universal grammar, the usage-based approach, input processing, processability theory, the declarative/procedural model, skill acquisition theory, interaction and complex dynamic systems (VanPatten, Keating and Wulff 2025). No theory covers the phenomenon as a whole; each explains one slice.

The status of the theories is a reading of separate literatures, not someone's single verdict [inference]; no direct comparison of the theories with one another was found. Behaviourism as a theory of language has been abandoned; its practical remainder, automatization through repetition, has been explained anew by skill acquisition theory, in which knowledge travels from declarative to procedural and automatic (DeKeyser and Suzuki 2025). Krashen's model has been rejected as a theory, but three of its intuitions are alive: comprehensible input is necessary, implicit knowledge differs from explicit knowledge, development is ordered (Lichtman and VanPatten 2021). The noticing hypothesis was softened by its own author: "for all practical purposes, attention is necessary" (Schmidt 2001). The usage-based approach gives the most concrete mechanism of interference from the native language, "learned attention": low salience of a form and blocking, both tuned by the native language, keep available input from becoming intake, and then "all the extra input in the world can sum to nothing" (N. Ellis 2006; N. Ellis and Wulff 2020). The declarative/procedural model is supported by a meta-analysis: vocabulary is linked to declarative memory, and grammar in adults is linked first to declarative memory and, with extensive experience, to procedural memory (Hamrick, Lum and Ullman 2018).

The dispute over whether a learned rule turns into fluent command has shifted towards "yes": automatized explicit knowledge predicted the acquisition of implicit knowledge in 100 advanced learners of Japanese (Suzuki and DeKeyser 2017), and over a semester explicit and implicit knowledge influenced each other reciprocally (Kim and Godfroid 2023). What remains unresolved is the question of measurement: what to count as implicit knowledge at all.

It is the absolute claims that failed, and in every case a weaker version survived. Matching instruction to a "style" (visual, auditory) is supported by practically nothing (Pashler et al. 2008). Children do not learn fast and effortlessly: in the first year in the environment, adolescents and adults progressed fastest, and children aged 3–5 were last on all tests (Snow and Hoefnagel-Höhle 1978). The "universal natural order" fell apart once there were seven native languages in the sample (Murakami and Alexopoulou 2016). The ability to learn a language does not switch off at puberty: the largest study places the decline at the end of adolescence and leaves its cause open, allowing that it is "an epiphenomenon of culture" (Hartshorne, Tenenbaum and Pinker 2018). For an adult the practical reading is this: a high functional level is attainable, being indistinguishable from a native speaker under strict scrutiny is practically unattested (Abrahamsson and Hyltenstam 2009), and analytical abilities partly compensate for age (DeKeyser 2000), which means that explicit study is rational for an adult.

The Soviet methodological tradition on which Armenian teachers rely arrived at some of these positions earlier than the West. The conscious-contrastive method of Shcherba's students required conscious rather than mechanical learning of rules, comparison with the native language, and a separation of receptive and productive command (Polyakova 2024). Galperin's theory of the stage-by-stage formation of mental actions is structurally close to skill acquisition theory and adds something it lacks, the quality of the initial orienting scheme [inference]; the dictionary of methodological terms notes that it works at the stage of skill formation, while attempts to apply it to free speech activity "proved to be of little success" (Azimov and Shchukin 2009). The sources disagree in assessing how closely the conscious-contrastive and conscious-practical methods are related: the dictionary sees a shared conception in them, Polyakova sees opposing camps. All of these are theoretical positions, not effect sizes.

What interventions show, and how far to trust them

This literature has its own yardstick: between groups, an effect of 0.40 counts as small, 0.70 as medium, 1.00 as large (Plonsky and Oswald 2014). Explicit instruction consistently beats implicit instruction on the tests that were used (Norris and Ortega 2000; Spada and Tomita 2010), but these tests mostly measure controlled knowledge; the most recent synthesis finds an overall effect of instruction of g = 1.06 and only "a minor difference" between explicit and implicit (Kang, Sok and Han 2019). Oral corrective feedback in the classroom gives a medium, durable effect, d = 0.74; prompts that make learners correct themselves are stronger than recasts, 0.83 against 0.53 (Lyster and Saito 2010). Two studies of the timing of corrective feedback found no gain in implicit knowledge in any group (Li, Li and Qian 2025): feedback by itself does not automatize.

Vocabulary is the most quantitatively worked-out area. Comprehension grows roughly linearly with the share of known words, with no threshold (Schmitt, Jiang and Grabe 2011); the working minimum is 95% coverage of a text, the comfortable level is 98%. For English this is 8–9 thousand word families for written text and 6–7 thousand for spoken text (Nation 2006); for informal listening 2–3 thousand families are enough at 95% coverage (van Zeeland and Schmitt 2013). The 98% threshold goes back to a single study of 66 students, and in its replication the results of the original study "could not be fully replicated" (Kremmel et al. 2023). What carries over to Armenian is the coverage percentages but not the family counts: in a suffixing language with compounding, the count depends on what is taken to be one word [inference]. Incidental learning from input is real but slow: 9–18% of target words immediately after a session; texts written for learners give g = 1.56 against 0.71 for texts written for native speakers (Webb, Uchihara and Yanagisawa 2023). Intentional learning gives about 60% immediately and 25–39% on a delayed test (Webb, Yanagisawa and Uchihara 2020).

Spacing and self-testing are the two techniques with the best evidence base in the whole science of learning, the only two out of ten with a high utility rating (Dunlosky et al. 2013). After the first correct recall, additional rereading gave nothing, additional testing gave a large effect, and the students' own predictions did not correlate with the result (Karpicke and Roediger 2008). In second-language vocabulary, expanding and equal intervals are statistically equivalent (Kim and Webb 2022): what is supported is "space it out", not a particular tuning of the algorithm. For grammar the picture is mixed, and a gap of 3.3 days beat a seven-day one (Suzuki 2017). A separate fact for the self-directed adult: students who used Anki showed a better result but "were reluctant to use the app and reported low enjoyment" (Seibert Hanson and Brown 2020).

Input alone is not enough for accuracy in production: after six years of a programme built on reading and listening, schoolchildren matched the control group in comprehension but not in written production (Lightbown et al. 2002); these are children, and there are no multi-year adult cohorts with testing. Grammatical skills are specific: comprehension practice builds comprehension, production practice builds production (DeKeyser 1997). Pronunciation instruction works, d = 0.68, but mainly on controlled production of specific sounds; for spontaneous speech the confidence interval may include zero (Saito and Plonsky 2019). Perceptual training improves perception, g = 0.67 against a control (Uchihara, Karas and Thomson 2025), and transfers weakly to production; the benefit of having many talkers specifically disappears after correction for bias, g = −0.04 (Zhang, Cheng and Zhang 2021). At the same time, a strong accent does not necessarily reduce intelligibility (Munro and Derwing 1995).

On a new script in adults the data are few, but they agree. Handwriting gave faster learning of Arabic letters and better transfer to untrained tasks than typing or viewing (42 adults, no replication) (Wiley and Rapp 2021); training the letter-sound mapping gave "striking benefits" over training the word-meaning mapping (Taylor, Davis and Rastle 2017). Familiar-looking letters bring with them the sound values of the native language (Hayes-Harb and Barrios 2021). No experiments on harm from transliteration in adults were found: the advice to "avoid it" is unproven.

All these numbers should be trusted with an adjustment. Replications in the field run at roughly one per 400 articles, and none are direct (Marsden et al. 2018); 94% of surveyed researchers admit to at least one questionable practice (Isbell et al. 2022); a reanalysis reduced the claimed effect of task-based teaching from d = 0.93 to 0.258 across three studies, and under the strictest criteria to −0.06 on one (Boers et al. 2021). The samples are mostly university students learning English. A reasonable expectation for any single intervention is a between-group effect of 0.4–0.7 on a test that matches the training, and less in spontaneous speech [inference]. Very reliable are spacing, retrieval practice, the link between vocabulary and comprehension, and the benefit of feedback; contested are the superiority of task-based teaching and the sufficiency of input; thin are the data on script, interleaving of forms, and language models in the role of tutor; branded methods have no data.

The native language explains up to a fifth of the outcome and sets the map of difficulties

The history of the question is two pendulum swings. The strong version of contrastive analysis (all difficulties are predicted by comparing the languages) failed: the predictions of four contrastive analyses on almost 2,500 Japanese schoolchildren turned out to be "inadequate predictors" (Whitman and Jackson 1972). The opposite conclusion of the 1970s, that the native language hardly matters, rested on children aged 5–8 and on counting only unambiguously classifiable errors, which is where the famous 4.7% comes from (Dulay and Burt 1974). Both extremes are wrong. The native language has a strong but selective influence, and not only through errors: also through avoidance of difficult constructions (Schachter 1974), speed, and under- and overuse.

The most reliable quantitative foothold is the state exam in Dutch. About 50,000 adults, 62 native languages: the native language explains from 9 to 22% of the total variance in the speaking score, that is, from 28 to 69% of what can be explained at all; the difference between the extreme groups is 72 points with a standard deviation of 37.8. About 80% of this effect is described by three measures of similarity: lexical (70%), phonological (44%) and morphological (31%) (Schepens, van Hout and Jaeger 2020). The direction is asymmetric: the score falls more because of features on which the target language is more complex than the native one (r = −0.67) than because of those on which it is simpler (r = −0.45) (Schepens, van der Slik and van Hout 2013); new sounds that the native language lacks get in the way, while the "extra" sounds of the native language do not. No such measure exists for Armenian as a target language; the transferability of the effect size to the Russian-Armenian pair is an assumption, while the direction is reproduced on three target languages.

Models of the mechanism differ in detail but agree on two things: the native language is the starting point, and the hardest part is functional morphology and the reassembly of how features are packaged into forms (Lardiere 2009; Slabakova 2013). Three results bear directly on Russian speakers. The first is an asset: attention to new verb endings is ordered by native language as "Chinese < English < Russian < Spanish" (N. Ellis and Sagarra 2011); a Russian speaker does have the habit of looking for grammar at the end of the word. The second is a caveat to the asset: what helps is a match in the way of marking, not the presence of the category as such (Sabourin, Stowe and de Haan 2006), and the speed of acquiring case depends on the reliability of the endings as a cue, not on the number of cases (Kempe and MacWhinney 1998). The third is a liability: a category that the native language lacks is the hardest to add. In a corpus of about 10,000 exam scripts, English articles are consistently at the bottom of the accuracy ranking for native speakers of Japanese, Korean, Russian and Turkish, and at the highest exam level morphemes with no equivalent in the native language are only as accurate as morphemes with an equivalent are at the lowest (Murakami and Alexopoulou 2016). The errors of native Russian speakers in English articles are systematic, not random (Ionin, Ko and Wexler 2004).

In phonology, hearing and pronunciation live on different time scales. Monolingual English speakers acquired three perceptual categories of stops "after only a few minutes of exposure" (Pisoni et al. 1982), but in a task that required holding the contrast in words, acquisition went differently and more slowly (Curtin, Goad and Pater 1998). The most direct data on Russian speakers come from a ten-year observation of 24 adult Slavic-speaking immigrants in Canada (18 of them Russian speakers): the aspiration of English /p/ grew from 0.032 to 0.054 seconds over the first eight months and after ten years stood at 0.057, against 0.082 for native speakers; age of arrival did not correlate with intelligibility, and the spread between individuals was "dramatic" (Munro, Derwing and Saito 2025). No studies were found of how speakers of a language with a two-way voicing system master a three-way contrast with aspiration.

For most Russian-speaking adults Armenian is a third language after English, and this changes the picture. The source of transfer can be both the native and the second language; data with Russian-speaking participants are consistent with property-by-property transfer from both languages, both helpful and harmful (Westergaard et al. 2017). The risk has been shown on Swedes: without English they did not violate German word order even once, with English they did so in 45% of cases (Bohnacker 2006). The opportunity has been shown on children: Danish-Russian bilinguals use English articles almost like native speakers, and Russian does not get in their way (Savelieva and Rodina 2022). Whether the English category of definiteness will transfer into Armenian in an adult with late English, nobody has studied [inference]. In large data sets the contribution of the second language is real but an order of magnitude weaker than that of the native language.

Should the method differ by native language? As regards content and emphasis, yes, and this is supported by a small but consistent set of experiments. Adding explicit explanation about the native language on top of identical instruction improved, in 50 English speakers, not only accuracy in the French imperfect but also the speed of real-time processing (McManus and Marsden 2017), and in the next experiment the accuracy of oral production both immediately and after six weeks (McManus and Marsden 2019). Contrastive analysis with translation beat two other conditions on all vocabulary tests (Laufer and Girsai 2008). The hardest features turned out to be not those unique to the new language but those that resemble the native language and work differently (Tolentino and Tokowicz 2014). There are no data in these studies on harm from using the native language; glosses in the native language give a small plus, g = 0.33 (Kim, Lee and Lee 2024). The caveats are substantial: the samples are small, the delay is no longer than six weeks, and not a single study was found that compares a course tailored to a particular native language with a universal course. As regards general principles, the method does not depend on the native language: from 78 to 91% of the variance in outcome is unrelated to it, and within one language group the spread is huge. The native language sets the task, but not the outcome.

A Russian speaker in Armenian: case comes for free, the article and aspiration do not

No published analysis of the errors of Russian speakers learning Armenian exists. There are three kinds of foothold: the report on RAU's experience; a fairly rich mirror-image literature on the errors of Armenians in Russian; and good descriptive linguistics of Armenian. For this reason almost all of the classification below is an inference from comparing structures, and it is tagged as such. The authority on the structure of Armenian here is the lecture materials of the Glottothèque and phonetic descriptions; Dum-Tragut's grammar, the main reference work, was not opened in this work, and everything known with reference to it is known second-hand.

The typological databases give two different pictures, and they should not be averaged [inference]. According to the WALS database, Russian matches Eastern Armenian on 54.7% of 117 shared features: as far away as from Georgian, Persian or Turkish, against 68–81% with European languages. According to the Grambank database the match is 73.1% of 186 features, no lower than that of Russian with English. Both numbers are our own count on open data (WALS; Grambank), with no weighting and no correction for dependent features, and several values in the databases for Armenian are disputable or erroneous, so the precision is a few points at best. The first picture is a better predictor of effort in syntax and phonology; the second explains why the nominal system feels familiar.

Phonetics: one new category and three habits to suppress

The inventory is mostly shared. The five vowels of Armenian match the Russian ones in quality, and Armenian vowels "do not affect the quality of preceding and following consonants" (Ovakimyan 2024). There is one new category with a high load, the three-way series of stops and affricates: 15 phonemes where Russian has 8. In Yerevan speech these are voiced aspirated, plain voiceless and aspirated; the contrast is maintained word-finally as well, which is typologically rare, and the script gives one letter per phoneme: բառ "word", պար "dance", փակ (Seyfarth et al. 2024; Dolatian 2022). An acoustic study of eight speakers raised in Yerevan found no ejective articulation in the plain voiceless series; the three categories are reliably separated by the intensity of voicing during closure and the duration of aspiration, in both positions (Seyfarth and Garellek 2018). A report of almost constant ejectivity word-finally in three speakers is known only from a search summary [unverified]; a later illustration reconciles the data as dependence on speaker and position. The WALS database codes this series as ejective, which is at odds with the acoustics.

Russian voiceless unaspirated stops map onto Armenian պ, տ, կ, so the new task is to add and, above all, to hear aspiration, which exists in Russian positionally and "is not noticed by native speakers" (Arustamyan 2003). The expected confusions are պ/փ, տ/թ, կ/ք, ծ/ց, ճ/չ [inference]: not a single perception or production experiment with Russian-speaking learners of Armenian was found. The RAU teacher names precisely ը, թ, փ, ք and ձ, ծ, ճ as difficult (Iskandaryan 2025); a second article from the same university adds that the aspirates are heard as plain stops and that ծ, ձ, ճ merge with ց and չ (Rostomyan 2026). Observations on perception, without an experiment, are in the 2013 dissertation (see above). Since the plain series is realized differently by different Yerevan speakers, it is logical to treat aspiration, not ejectivity, as the target feature [inference].

The rest is not switching on something exotic but switching off three automatic Russian rules. Transfer of vowel reduction is attested: բոլորը is pronounced as "балоры" (roughly "balory", with the unstressed vowels reduced the Russian way). Transfer of palatalization is attested: տ, դ, լ are softened before front vowels. Armenian phoneticians described the same by ear as early as 1971 and 2006 (Տիգրան, մատիտ) and noted this softening above all in Russian speakers (Karagyulyan 1971; Zakaryan 2006). Transfer of stress is attested as a class of errors: in Armenian it falls on the last full vowel (Iskandaryan 2025). Devoicing of final voiced consonants by Russian speakers is only a prediction [inference], resting on a mirror-image fact: Armenians speaking Russian keep the voicing in "ряд" ("row") and "мороз" ("frost") and do not distinguish hard and soft consonants, and "the zones of maximal interference are the zones of cross-linguistic difference" (Arustamyan 2003). The profile is asymmetric [inference]: for an Armenian in Russian the hard part is building a new category; for a Russian in Armenian it is mostly suppressing three habits and building one. Suppression errors are persistent but rarely interfere with comprehension, with one caveat: reduction of unstressed vowels erases case endings, which carry grammar. The mnemonic from a methodological article, "ր is a soft r, ռ is a hard one", is phonetically inaccurate (the contrast is tap versus trill) and would push towards palatalization [inference].

Script: transparent, with one hidden depth

Teachers call the alphabet the main entry barrier, and in all traditions a separate introductory course is devoted to it; at RAU the "psychological barrier" is lowered by reading proper names and words that coincide in the two languages (Iskandaryan 2025). The reformed orthography adopted in Armenia is close to the principle of "one sound, one letter": 36 sounds, 39 letters (Ovakimyan 2024). The exceptions form a short closed list and sit in the most frequent words [inference]: initial ե is read [je], initial ո is read [vo], և is [jev] or [ev] depending on position; the most frequent word form in the corpus is է "is".

The real depth of the system is the unwritten schwa. The spelling shows long strings of consonants that in speech are predictably broken up by a vowel: դնել [də.nel] "to put", մկրտել [mə.kəɾ.tel] "to baptize", and the "location of schwa depends on a lot of factors (morphology, sonority, directional syllabification)" (Dolatian 2022). A Russian speaker, tolerant of consonant clusters, will insert it less often than required [inference]. In the opposite direction, from sound to letter, the three-way series turns dictation into a perception task: someone who cannot hear տ and թ cannot write them down, so ear training and spelling are one and the same thing here [inference]. The lists of visually similar letters that are usually given are not confirmed by any study [unverified].

No studies were found of the transparency of Armenian orthography, of learning to read in Armenian, or of the time needed to master the alphabet. Adjacent data say that the skill of letter-sound decoding itself transfers between alphabets in a literate adult: readers with a Cyrillic and with a Latin native script did not differ significantly in English (Mansuri et al. 2025), so what has to be learned is the correspondences, not the principle [inference]. The US Foreign Service Institute (FSI) nevertheless warns from experience: bringing the processing of a completely foreign script to a level of automaticity sufficient for concentrating on meaning takes an adult much longer than reading researchers had assumed (Jackson and Kaplan 1999).

Morphology: an easy nominal half and a hard verbal half

Nominal morphology is the easy half. There is no gender, adjectives do not agree, the order of suffixes is number, case, determiner, and allomorphy is small and phonological: plural -եր after monosyllables and -ներ after polysyllables, definite -ը after a consonant and -ն after a vowel (Dolatian 2022). A Russian speaker already has case as a category, and this is the largest advantage over an English-speaking learner [inference]. The transfer, however, is conceptual, not formal: in Armenian the nominative coincides with the accusative and the genitive with the dative, while the ablative and the locative are separate cases, so a Russian speaker will look for an accusative ending and will over-distinguish the genitive from the dative [inference]. The teacher reports that the very shift from the fusional type to the agglutinative one is felt as a difficulty, and the use of the singular for a generalized plurality (Նա ծիրան շատ է սիրում, literally "he apricot very much loves") is separately named as difficult (Iskandaryan 2025). Irregularity sits on frequency: special stems in words such as ընկեր "friend" and մայր "mother", a suppletive aorist for գալ "to come" (Dolatian 2022); for this reason "regular agglutination" reaches the beginner late [inference].

Definiteness is probably the largest grammatical cost [inference]. A Russian speaker has no category to push off from, and the Armenian article is not simply "the". It is obligatory with demonstrative and possessive pronouns and agrees with the person of the possessor (իմ նկարս "my picture", քո նկարդ "your picture"), it is obligatory on the head of a genitive construction, and it works as a nominalizer, but it does not attach to genitive, ablative, instrumental and locative forms, and places and goals of motion go without an article and without case: Վարդանը Երևան է "Vardan is in Yerevan" (Hodgson 2022). Object marking is also tied to definiteness: an inanimate or non-specific object stands in the zero form, an animate specific one in the dative, hence Ոստիկան եմ փնտրում "I am looking for a policeman (any one)" as against the form with the dative for a particular person. The Russian animate accusative gives a handhold here, but the Armenian trigger also includes specificity. The expected errors are omission of the article on subjects and definite objects and overuse on places; there are no learner data that would confirm this.

The verb is the second hard zone. The present indicative is analytic: a participle in -ում plus an auxiliary (գրում եմ "I write"), while the old synthetic present has become the subjunctive; three forms compete for the future: կգրեմ, գրելու եմ, պիտի գրեմ; there is no progressive (Khurshudyan 2022). The perfect in Eastern Armenian has an evidential reading: inference, hearsay, surprise (Donabédian-Demopoulos 2022). The copula is obligatory, and at RAU it is explained to Russian speakers through the English "I am", because Russian has no copula in the present tense (Iskandaryan 2025). A Russian speaker conceptually has the perfective/imperfective distinction in the past; what turns out to be new is the opposition of aorist and perfect [inference]. It is telling that in the same RAU article, under the heading "pluperfect tense", the examples given are with the aorist and the perfect: this area is confusing even in a teacher's exposition [inference]. It follows from the mirror-image literature that Russian aspect pairs and verbs of motion are hard for Armenians, which means that a Russian speaker will not have to learn these subsystems in Armenian; the articles themselves, however, were seen only in search summaries [unverified].

Syntax: not "SOV", but focus before the verb

The formula "Armenian is an SOV language" is not supported by the data, and it should be removed from explanations for Russian speakers. In a corpus sample of 570 transitive clauses with an overt subject, object and verb, SVO order made up 78.9% and SOV 16.3%; with definite objects, 79.1% against 20.9% (Faghiri and Samvelian 2020). The authors state a limitation: the sample is unbalanced and built on two verbs with predominantly definite objects. A continuation of this work, according to its abstract, shows that the position of the object is determined mainly by definiteness; beyond the abstract this is [unverified]. The lecture materials put it directly: word order in Armenian is determined by information structure more than by grammatical relations (Hodgson and Khurshudyan 2022). The precise short formulation is this [inference]: article-less objects and other focal material stand immediately before the verb and carry the auxiliary; definite objects usually follow the verb. For this reason conclusions about the early acquisition of verb-final order obtained on other languages apply only weakly to Armenian.

What is new lies below and above the level of the clause. The auxiliary is a clitic that attaches to the first constituent of the predicate or to the focus: ԴԱՎԻԹՆ է Նունեին գիրք նվիրել "it was DAVIT who gave Nune a book" as against the neutral Դավիթը Նունեին ԳԻՐՔ է նվիրել; it cannot attach to postverbal elements (Hodgson and Khurshudyan 2022). The teacher formulates the same thing in Russian: the logically important component "attracts the auxiliary to itself" (Iskandaryan 2025). This is the syntactic twin of the article problem [inference]: a small unstressed element with no Russian counterpart whose position encodes a meaning that Russian encodes with intonation and word order. The expected behaviour of a Russian speaker is to omit it or to glue it rigidly after the participle in every context; there are no studies that would confirm this error pattern. The noun phrase is strictly head-final: the genitive before the noun, postpositions instead of prepositions, prenominal participial modifiers. At RAU this is made visible by translating with Armenian word order: Արմենի տունը is rendered as "Армена дом" ("Armen's house" with the possessor first, where neutral Russian puts it after the noun). This area is completely regular, that is, it requires drilling rather than understanding [inference].

Vocabulary: an opaque core and a hidden asset in calques

Armenian is a separate branch of the Indo-European family, and its basic vocabulary is opaque to a Russian speaker; the teacher says directly that it "differs greatly from the Russian, which makes memorizing words harder", and that the genetic kinship "can in many cases be revealed only through careful linguistic analysis" (Iskandaryan 2025). Some of the pairs in that article's teaching list are not standard cognate sets, and the list was not checked against an etymological dictionary [unverified]. On the first thousand words a Russian speaker gets almost no free vocabulary, and the bulk of the work is vocabulary, not grammar [inference].

Transparency comes from two contact layers and is inversely proportional to the prestige of the register [inference]. Russian had a stronger influence on the written language and the press through the bilingual education of recent generations (Donabédian-Demopoulos 2022), while Russian loanwords live in colloquial Yerevan speech; the literary standard replaces many of them with its own coinages. A rough indicator of the size of the layer: in the English Wiktionary, as of 13 October 2026, 1,140 Armenian entries are tagged as borrowings from Russian and 194 as calques from Russian, out of 19,278 lemmas (borrowings category; calques category). This is the coverage of the editors, not a sample of the lexicon. The hidden asset is calques and transparent compounding: teachers use word nests such as օր "day", այսօր "today", կեսօր "noon". How far Armenian compounds parallel Russian ones morpheme by morpheme is not measured in the sources [unverified]. A lexicostatistical distance for the Armenian-Russian pair could not be obtained, and it is lexical similarity that is the main carrier of the native-language effect in large data sets. This is the most important unmeasured parameter of the whole picture.

Calques have a reverse side. Armenian studies of Russianisms list the same constructions carried over from Russian again and again: ինձ մոտ for «у меня» ("I have", "with me": Ինձ մոտ դա չի ստացվում instead of Դա ինձ չի հաջողվում, "it isn't working out for me"), փողոցի վրա for «на улице» ("in the street") instead of փողոցում, Մեզ անհրաժեշտ է իմանալ ("we need to know") with a dative instead of a nominative subject, an infinitive after որպեսզի ("so that") where Armenian requires a finite verb (Khachatryan 2020; Muradyan 2006). These works give no counts; they are normative lists. They were compiled from the speech of Armenians who know Russian, so for a Russian-speaking learner the risk is double [inference]: the learner will build the same forms out of native habit and hear them all around in Yerevan. No general frequency dictionary of Eastern Armenian was found in the Armenian literature; there is a 1968 thematic list of schoolchildren's spoken vocabulary, 2,602 words (Kazaryan 1968a). The same author wrote that the Armenian textbooks for non-Armenian schools of 1963–1965 were compiled from the authors' intuition and own experience, with unsimplified texts and rare words (Kazaryan 1968b).

Summary map

All rows of the table are an inference from comparing structures, except for the points attested by the RAU teacher: reduction, palatalization, stress, the difficulty of the schwa and of the aspirates, the generalized singular, and the explanation of the copula through English.

The conjectured order by cost for this learner, from greater to lesser [inference]: volume of vocabulary; shortage of input and of necessity; the article together with the copula and the auxiliary; perception of aspiration and the three habits to suppress; the correspondences of tenses and moods; the script as a bounded cost paid up front; nominal morphology as a net saving compared with what a Russian speaker expects.

Analogues, environment and resources: what to expect and what to build from

Analogue languages set realistic expectations

No studies of adults acquiring Armenian or Georgian were found, so expectations about timing are taken from evaluated programmes for other small or case-marking languages. The Gaelic Ùlpan is 216 hours, 144 lessons: after about 160 hours 69% of new learners reached A2 in oral production, and the levels, judging by the titles of the report's tables, are self-assessed; the course worked worse for complete beginners than for those who had prior listening experience, and the average time to completion was 3.43 years (MacLeod et al. 2015). In Finnish integration training in 2013–2016, fewer than 35% of those who completed it reached level B1, and the ministry "lacks sufficient information on the effect of training on language skills" (National Audit Office of Finland 2018). In Israel each month of ulpan is associated with a gain of 0.08 scale units in self-assessed fluency, that is, about a year to move to the next step (Chiswick, Rebhun and Beider 2016). Both of the first two languages use the Latin alphabet; Armenian adds on top the cost of automatizing the script [inference].

Laboratory studies of adults mastering Russian case give the mechanism. Generalization turned out to be "far worse than would be expected if learners had extracted the underlying rules"; inflectional systems require considerable item-by-item learning (Kempe and Brooks 2008). A larger training vocabulary improved generalization only in participants scoring above the median on a non-verbal test (Brooks, Kempe and Sionov 2006); age affected pattern-based processes but not memory-based ones (Kempe, Brooks and Kharkhurin 2010). Explicit deductive instruction is better for simple categorical rules (DeKeyser 1995). Since Armenian declension is more regular and more agglutinative than Russian declension, the share of rule should be larger and the item-by-item load smaller and concentrated on the irregular minority of frequent words [inference]. The information that in Turkish and Korean the main error in case markers is omission rather than wrong choice is taken from search summaries [unverified]; if it is so, for Armenian it predicts omissions of the article and of markers in speech.

The experience of the Foreign Service Institute is the only institutional practice with a stated success rate: about two thirds of students reach the goal. Its lessons in its own formulations: time on task and intensity are decisive, and shortening the programmes "has not worked"; groups of four people or fewer; focused practice, including drills, is needed by almost everyone; explicit knowledge of grammatical concepts speeds up progress; adults need a lot of "easy" reading; in-country immersion gives the greatest return above level two on the ILR scale; ordinary conversation, especially multi-party conversation, is one of the hardest skills (Jackson and Kaplan 1999). The frame for a weekly plan is given by the four strands scheme: roughly a quarter of the time each on meaning-focused input, meaning-focused output, language-focused learning and fluency development, although the author himself calls the equal division an "arbitrary decision" (Nation 2007). There is a real contradiction here: according to data the author himself cites, intentional learning gives several times more words per hour than reading a graded reader, which argues for a larger share of intentional work at the very beginning, yet he does not support such a shift.

Environment: in Yerevan, Russian is always at hand

The main obstacle for a Russian speaker in Armenia appears to be ecological, not structural [inference]. What is known about repatriates from Russia after February 2022 comes from the abstract of a conference paper: since schools and kindergartens offer classes in Russian and most employees of state institutions speak Russian, not knowing Armenian "has not posed practical issues in everyday life" (Tokmatcev 2023); these are interviews without numbers, and about ethnic Armenians, not about relocants in general. The switching mechanism has been measured in Finland: 95% of 358 surveyed native speakers of Finnish report that they switch a conversation with a non-native speaker to English, for reasons of efficiency, politeness and a wish to help, believing that the interlocutor is glad of it (Heikkola et al. 2024). Carrying this over to Yerevan with Russian in place of English is an inference, not data; there is no survey measuring switching in Armenia. The Israeli analogue (a self-sufficient Russian-speaking environment as a restraining factor for Hebrew) is known from a search summary [unverified]. The share of Russian-speaking relocants learning the language of the host country was not found for either Armenia or Georgia: the report on Russians in Georgia contains no questions about language at all (CRRC-Georgia 2023). The Russian-language scholarly literature on this topic was not reviewed, and that is the most likely place where such data may be.

The second feature of the environment is the gap between the standard and living speech. The RAU teacher calls it "one of the significant difficulties", and starts the course itself with colloquial formulas, for example with ոնց instead of the literary ինչպես (Iskandaryan 2025). No comparative studies of the order in which the standard and the colloquial language are taught were found for any language; the integrated approach for Arabic is argued for and described, but it has almost no controlled outcome data, and this is known from publishers' descriptions [unverified].

The third is attrition. In a study of self-directed learning with software in the workplace, out of 150 volunteers one person completed the full 200-hour protocol, and 51% never logged in at all (Nielson 2011). The task of a self-directed adult is persistence before method [inference].

There are enough resources to engineer a course, but all of them are about the literary standard

Eastern Armenian is better supplied with data than with pedagogy. The Eastern Armenian National Corpus contains about 110 million tokens, including a spoken subcorpus of about 3.5 million, with English translations for roughly 85% of tokens; the analyzer does not recognize 7.5% of tokens, and most of those are in spoken language (Khurshudyan et al. 2022). The repository of the openly licensed morphological analyzer contains a frequency list of the corpus's word forms (uniparser-grammar-eastern-armenian); no lemmatized frequency dictionary and no graded list for learners were found. There are two syntactically annotated corpora (UD Armenian-ArmTDP), a parallel Armenian-Russian corpus of 2.4 million tokens inside the Russian National Corpus (Khurshudyan 2022), and a transcription for almost every one of the 19,278 Armenian lemmas of the English Wiktionary, which makes it a checkable source for the place of the schwa.

The quality limitations are substantial. In a 2026 speech recognition benchmark the best system gives a strict word error rate of 13.80% on the combined set, and the median error on film dialogue is 61.67% (ArmBench-ASR); the leaderboard figures changed between visits. In the language model benchmark the best average score is 0.635 (ArmBench-LLM). Two conclusions follow [inference]: automatic checking of aspiration should not be trusted without a separate test on recordings of native speakers, and model-generated examples and explanations need to be checked against the analyzer, the corpus and a native speaker. A meta-analysis of feedback from speech recognition gives g = 0.69, but only for English (Ngo, Chen and Lai 2024). All the tools listed are aimed at the literary standard; colloquial Yerevan speech, which is what a learner living in the city hears, is covered worst of all.

Conclusion

The state of affairs turned out not to be what the question presupposes. There is a science, but not about Armenian; there are methods, but without measurements; the description of the language is good, but not one prediction about the Russian-speaking learner has been tested on learners. The emptiness, however, is conveniently shaped: it is almost entirely empirical. There are enough theoretical footholds, the tools for assembling a course are open, and the missing link, learner data, is produced by the learning process itself if measurement is built into it. A documented, research-based method for a small language with even modest tracking of outcomes has, as far as could be established, no clear precedent; the evaluation of the Gaelic Ùlpan shows how methods without such tracking end up: learners were not aware of what progress they were making.

The second thing that changed in the course of the work is the idea of where the difficulty lies for a Russian speaker. Not in the cases and not in the word order, which are what usually frighten people, and not even in the alphabet, whose cost is bounded. It lies in three small unstressed things that Russian does not have (the article, the copula, the auxiliary), in one feature that the Russian ear does not treat as distinctive (aspiration), in a vocabulary without footholds, and in a city where the language is not required. A solitary learner keeping a log of their own errors turns out here to be not a consumer of the theory but its first source of data.

Candidate principles: fifteen hypotheses and ways to test them

Every proposition below is a hypothesis, not a recommendation. For each one it is stated what it stands on and what kind of support that is, and what kind of study could confirm or refute it. The first seven do not depend on the native language; the next eight concern those whose first language is Russian.

General hypotheses, for any native language

H1. Letters are consolidated fastest by handwriting and by training the letter-sound mapping, and the closed list of irregular frequent spellings and the schwa rule are best given explicitly in the first weeks. Support: verified experiments on other scripts, namely the handwriting advantage in 42 adults (Wiley and Rapp 2021, no replication) and the advantage of print-to-sound training (Taylor, Davis and Rastle 2017); the description of Armenian orthography and the unwritten schwa (Dolatian 2022). The transfer to Armenian and the claim about harm from transliteration are an inference; there are no studies of the Armenian script in adults. Test: randomize beginners to handwriting, typing and a transliteration crutch; measure the speed and accuracy of reading aloud, including words with an unwritten schwa, after one, four and twelve weeks.

H2. Vocabulary is best built from frequency, by intentional learning with spacing and retrieval practice, and at the very beginning it should get more than a quarter of the time. The support is verified: the linear relation between coverage and comprehension (Schmitt, Jiang and Grabe 2011), spacing and self-testing (Dunlosky et al. 2013; Kim and Webb 2022), the slowness of incidental learning (Webb, Uchihara and Yanagisawa 2023). The frequency principle has no randomized trial of its own and follows from the arithmetic of coverage; the shift of proportions at the start contradicts the recommendation of the author of the four strands scheme; the English word-family numbers do not carry over to Armenian. Test: lemmatize the corpus frequency list, measure coverage of the spoken subcorpus by frequency band, and compare a "frequency order" group and a "thematic order" group on comprehension of Yerevan speech after three months.

H3. Functional morphology is acquired through a short explicit rule followed by practice in the modality that is needed, on many different words; paradigm tables alone are not sufficient. The support is verified: the specificity of skill (DeKeyser 1997), the advantage of explicit instruction for categorical rules (DeKeyser 1995), item-by-item learning of Russian case by adults (Kempe and Brooks 2008), functional morphology as the bottleneck (Slabakova 2013). Inference by analogy: in the more regular Armenian declension the share of rule is larger, and the item-by-item load is concentrated on irregular frequent words. Test: compare "rule plus varied words" with "rule plus table" on Armenian case forms, measuring transfer to new nouns in oral production after a week and a month.

H4. Comprehensible input is necessary but not sufficient, and the bottleneck for Armenian is the supply of graded texts and audio, not the method. The support is verified: a multi-year comprehension programme with no gain in written production (Lightbown et al. 2002, children); the twofold difference between texts for learners and texts for native speakers (Webb, Uchihara and Yanagisawa 2023); the absence of Armenian materials above B1 (Kagirova 2026, from the abstract). Inference: texts generated by a language model within frequency bands and checked by a native speaker are able to close the gap; there are no results of such an approach with learners of a low-resource language. Test: compare the gain in vocabulary and comprehension from reading such texts and unadapted texts with glosses, at equal time, with a delayed test.

H5. Colloquial Yerevan Armenian and the literary language are best run in parallel, split by modality: listening and speaking on colloquial forms, reading and writing on the standard. The support is weak, and this should be left visible: the teacher's testimony about the gap (Iskandaryan 2025), an approach for Arabic that is argued for but untested, complaints of learners of Finnish from a non-primary source. There are no comparative studies of the order of teaching for any language. Test: two groups with identical hours, "standard only" and "standard plus colloquial forms from day one"; the outcome is comprehension of recordings of spontaneous Yerevan speech and reading of notices and official texts after six months.

H6. Living in the country does not by itself give a beginner input: input and necessity have to be constructed, and without an external structure most self-directed learners drop out in the first ten hours. The support is verified: switching to a shared language in Finland (Heikkola et al. 2024), the return on immersion above level two (Jackson and Kaplan 1999), attrition in self-directed learning (Nielson 2011), the abstract about repatriates (Tokmatcev 2023). The transfer to Yerevan is an inference; the frequency of switching there has not been measured. Test: first a survey of relocants on how often interlocutors switch to Russian and whether they are learning the language; then a comparison of learners with and without pre-arranged "Armenian" situations (a tutor, particular shops, acquaintances) on hours of real communication in Armenian and gain in spoken language.

H7. A method with measurement built into it outperforms a method without it and at the same time produces the data the science lacks. The support is verified: the argument that group means do not predict the trajectory of an individual (Lowie and Verspoor 2019); the evaluations of the Finnish and Gaelic programmes, where the absence of measurement is named as a problem. Inference: an error log kept against a predefined list and periodic recordings of speech are a workable form of such measurement. Here the test coincides with the product: an annotated corpus of errors by learners of Armenian, which does not currently exist, with the national corpus in the role of the reference.

Hypotheses for those whose first language is Russian

H8. Aspiration is the main perceptual risk, and discrimination training on minimal pairs with feedback is most effective in the first months and only together with consolidating the contrast in words through dictation. The support is verified: the acoustics of the Yerevan contrast (Seyfarth and Garellek 2018); the ten-year trajectory of aspiration in Slavic speakers, with growth in the first six to eight months (Munro, Derwing and Saito 2025); the gap between hearing and the lexicon (Curtin, Goad and Pater 1998); the effect of perceptual training (Uchihara, Karas and Thomson 2025). Inference by analogy: that the Russian speaker's task in Armenian is the same, only with the voiceless stops split into two categories that have to be kept apart; that the target feature is aspiration, not ejectivity; that it is worth starting with low stimulus variability (the last point rests on a study known from a summary). Test: an identification and discrimination experiment on five pairs with Russian-speaking listeners before and after training, with a measurement of voice onset time in their own speech and with a dictation test; no such study exists for any pair of languages of this type.

H9. Reduction, palatalization and final devoicing require explicit explanation in Russian and separate practice; reduction deserves priority because it erases case endings. Support: reduction and palatalization are attested by the teacher (Iskandaryan 2025), devoicing is an inference from mirror-image data on Armenians in Russian (Arustamyan 2003); pronunciation instruction acts on controlled production of specific features (Saito and Plonsky 2019). The claim about the priority of reduction is an inference. Test: acoustic analysis of recordings of Russian-speaking learners at three stages of the course (quality of unstressed vowels, voicing of final consonants, secondary articulation before ի and ե) and a test of how well native speakers understand their case forms.

H10. The definite article is the longest-lasting grammatical difficulty for a Russian speaker, and it will show up above all as omissions in speech. The support is verified: the structure of the article (Hodgson 2022); the persistence of errors in English articles among native speakers of Russian up to the highest levels (Murakami and Alexopoulou 2016); their systematicity (Ionin, Ko and Wexler 2004). Inference: the transfer from the English article to the Armenian suffixed one; there are no data at all on the acquisition of a postposed article by native speakers of article-less languages. The prediction about omissions rests on unverified information about Turkish and Korean. Test: in an error corpus, compare the share of omissions and of overuses of the article in Russian-speaking and English-speaking learners of the same level, with three kinds of evidence of transfer by Jarvis's standard (Jarvis 2000).

H11. The copula and the auxiliary are acquired better if the rule for their position (attachment to the focus) is introduced in the first weeks on question-answer pairs, rather than after the order "participle plus auxiliary" has been consolidated. The support is verified: the description of the rule (Hodgson and Khurshudyan 2022) and its formulation by the teacher for Russian speakers (Iskandaryan 2025); the corpus predominance of SVO, from which it follows that the "SOV" formula is misleading (Faghiri and Samvelian 2020). The predicted error pattern (omission and rigid gluing) is an inference. Test: elicited production of answers to questions targeting different constituents of the sentence in two groups with early and late introduction of the rule; the outcome is the share of answers with the auxiliary on the focus.

H12. Case is a saving for a Russian speaker, but contrastive explanation pays off precisely on the partial matches: the syncretism of genitive and dative, the ablative instead of prepositional constructions, the locative, the dative of the object. The support is verified: the attention of native speakers of Russian to endings (N. Ellis and Sagarra 2011); the decisive role of similarity of marking rather than presence of the category (Sabourin, Stowe and de Haan 2006); the greatest difficulty with what is similar but works differently (Tolentino and Tokowicz 2014). The application to Armenian is an inference; no studies with readable results were found of Russian speakers mastering the case system of another language in comparison with native speakers of caseless languages. Test: compare the speed and accuracy of case marking in Russian-speaking and English-speaking beginners under identical instruction, separately for matching, partially matching and new functions.

H13. Targeted explicit comparison with Russian speeds up acquisition where the languages diverge in the means of expression, and where Russian has no category at all, English can serve as the support, with the risk that it will also bring false transfer. The support is verified: the experiments with added information about the native language (McManus and Marsden 2017; 2019), the data on a third language in Russian-speaking participants (Westergaard et al. 2017; Savelieva and Rodina 2022), the harm of transfer from the second language (Bohnacker 2006); RAU's practice of explaining the copula through English. The samples are small, the delay is up to six weeks; the configuration "Russian plus English, then Armenian" has not been studied. Test: a randomized comparison of three conditions on the aorist and the perfect: explanation on Armenian material only, with comparison to Russian aspect, and with comparison to Russian and English; the outcomes are accuracy, processing speed and oral production after six weeks and six months.

H14. The basic vocabulary does not help a Russian speaker, but calques and transparent compounding give an advantage that grows after the first few hundred roots, and colloquial Russianisms are useful only with a register label. Support: the teacher's testimony about the opacity of the core and about the use of word-formation nests (Iskandaryan 2025); lexical similarity as the main carrier of the native-language effect (Schepens, van Hout and Jaeger 2020); the advantage of cognates in learning (de Groot and Keijzer 2000). The calque advantage itself is an inference, not measured by any source; there is no lexical distance for the pair. A test in two steps: count the share of borrowings and calques from Russian by frequency band, separately for the written and the spoken corpus; then compare the speed of learning and the retention of calqued and non-calqued compounds in Russian-speaking and English-speaking learners.

H15. The American estimate of a thousand-plus class hours is an upper reference point, not a forecast for a Russian speaker, and an estimate of one's own can be obtained only by accumulating data. The support is verified: all three versions of the figure (1,012 hours on the current page, 1,100 in the archived copy and in the 2024 article, the absence of Armenian from the 1999 table) were measured on selected native speakers of English (state.gov; Jackson and Kaplan 1999); the asymmetry by which adding complexity is harder than losing it (Schepens, van der Slik and van Hout 2013). The direction of the adjustment for a Russian speaker is an inference: less morphological complexity to add, but the same opaque vocabulary and the same new script. Test: keep a record of hours and of level by the 2024 state descriptors in a cohort of Russian-speaking learners; the very first hundred such records will be the only data in the world on time to level for this pair of languages.