The first step is a word-wise analysis of the verse by the libEscansión class PlayLine, whose practical usage is explained below. It consists of dividing the words into syllables and determining whether a particular word carries prosodic stress. The former task is carried out by the library Fonemas [Sanz-Lázaro 2021a], whilst the latter resorts to Stanza [Qi et al. 2020] PoS-tagger to determine the grammatical category and whether it is stressed. This class has two public attributes: a string containing the verse and a list of its words, each of whose elements is, in turn, another list with the corresponding word syllables, which are marked if they carry metric stress.

Even though a phonological transcription is not strictly necessary to analyse a verse, it is convenient for diverse reasons. Firstly, using IPA notation with a single Unicode character for each phoneme expedites the straightforward handling of syllabic division, as we can treat each character equally without exceptions[2]. Spanish digraphs (ch, rr, ll, qu, gu) or multiple phonemes represented by a single grapheme (e.g., x) are no longer required. Second, the output allows an out-of-the-box rhyme analysis as the relation phoneme-character is biunivocal, which spares a special treatment for ambiguous graphemes (b, v, c, z, g). In this step, the text is not only transcribed but also preprocessed. Stranded letters are transliterated to their spelling (e.g., B would become be) as well as non-Spanish diacritics (e.g., Catalan què would become ké); other potential sources of errors are also handled here (e.g., typographic ligatures). Finally, it allows a straightforward treatment of vowel clusters, as syllabic and non-syllabic vocalic allophones do not share a common representation.

The syllabification proper is carried out by the independent library Silabeador, called by the transcriber. First, the programme looks for non-transliterated Latin words and tries to transliterate them according to the Spanish orthographic rules. Contrary to other foreign words, Latin words are not phonologically adapted and tend to respect not only stress positions contradicting Spanish rules but also the original syllabic division, which may also differ from Spanish due to the disparate sets of diphthongs of both languages. Then, the programme marks forced syllabic divisions where an exception should occur according to an external file. This file consists of two columns with regular expression substitutions; the first field contains a matching pattern, and the second corresponds to its substitution, including an underscore where a syllabic division must be forced. These phonological exceptions include ubiquitous words whose Castillian pronunciatiation differs from the orthographic hyphenation. To name some extremely common in any kind of speech, this rule matches the hiatuses of verbs in -iar, -uar, -uir [Quilis 2019], -eir, and -oir [Ríos 1999], adjectives in -uoso [Quilis 2019], some particular nouns [Navarro 1957], and other words [Canellada and Madsen 1987]. Adjectives in -mente are also exceptional since they carry secondary stress [Quilis 2013], which Spanish phonology otherwise lacks. They are marked too in the morpheme, and the syllabification module receives the lexeme alone; finally, both parts are rejoined after their separate treatment.

There are ambiguous words, such as juez, suave or fiel [Domínguez 2012], in which the diphthong is often broken in Spanish Golden Age verses. They might be typographically marked with a diaeresis (ï or ü) by the editor of the source, in which case they are parsed as usual, but this may not be the case. Therefore, these ambiguous clusters are marked as a diaresis, although their sinaeresis will be given priority later.

The resulting text is divided in its syllabic nuclei, arranging the diphthongs except for those explicitly separated in the previous step. Non-syllabic vowels – the non-nuclear sounds in a vocalic glide – are marked as such by substituting them for the character representing the corresponding allophone. The outcome is a list of syllabic nuclei and free consonants. Finally, those free consonants are assigned to the proper syllable on coda or onset according to Spanish phonological rules [Quilis 2019]. The output is a list of phonological transcribed syllables and the prosodic stress on the due ones.

Word stress does not correspond to metric stress exactly[3] as some word classes are considered unstressed regardless of their pronunciation in isolation [Quilis 2013]. The class VerseMetre of libEscansión determines which words carry metric stress. It takes a verse as an argument, cleans it, and divides its words into syllables using the library Fonemas. It also uses Stanza [Qi et al. 2020] to PoS-tag each word. After this, it decides whether a particular word carries stress according to its PoS-tag, but also considering the relation to neighbouring words when forming a compound as it may deprive from their prosodic stress to otherwise stressed words (e.g. Santo Tomás /san.to toˈmas/ but el santo padre /el ˈsan.to ˈpa.dɾe/ or cuarenta /kwaˈɾen.ta/ but cuarenta y dos /kwa.ɾen.tai̯ˈdos/), in a similar fashion as it occurs, for example, with English numerals.

PoS-tagging may produce consistent errors or not be detailed enough. It can be bypassed in such situations by marking the words according to lists of stressed or unstressed occurrences. Undetected errors, however, are still possible. The greatest difficulty is detecting when an unstressed word should not carry prosodic stress as it involves several open word classes, whilst unstressed words belong to close ones. Therefore, a relatively high proportion of wrongly tagged words has no impact on the outcome as long as the tag represents an equivalent class in terms of stress.

We have tested the syllabification accuracy against the 106,362 words corpus EDFU [Rosa and Pérez 2020]. Using an American configuration as the corpus does, that is, without non-orthographic hiatuses and the group tl as an indivisible onset (trans-a-tlán-tico instead of trans-at-lán-ti-co), the syllabification module achieves 99.92% accuracy. This improves Agirrezabals's [Agirrezabal, Alegría and Hulden 2017] 98.06% and Navarro-Colorado's [Navarro-Colorado 2015] 98.35%. However, at first sight, it lies behind 99.99% of Rantanplan [Rosa et al. 2020a]. That is, disregarding errors like improper diphthongs in the corpus, such as amambaiense a-mam-baien-se instead of a-mam-ba-ien-se. All in all, 86 words differ from the corpus. From those, after a visual inspection, 37 differences are wrongly syllabified in the corpus, 41 have alternative yet acceptable differing syllabifications, and our programme wrongly syllabifies eight words Table 2.

The words penthouse and ouija do not obey the Spanish rules and are syllabified by approximating the original phonology, which would require to treat them as an exception. Since this would not be productive, we have opted to accept this kind of error. When considering only these eight proper errors, the accuracy rises to 99.99%.

To test the accuracy[4], we downloaded the corpus ADSO 100 [Navarro-Colorado 2017] using the tool Averell [Díaz et al. 2021] to compare our results with previous studies, which is the gold standard they used. ADSO 100 is the only manually annotated corpus of its size with a reported inter-annotator agreement. It comprises 1,404 verses.[5]. Our hardware was an Intel® Core™ i5-3470 CPU @ 3.20GHz, 16 GB of DDR4 RAM and a GPU NVIDIA GeForce GTX 1060 with 6GB. The PoS-tagging module ran on the GPU. For the syllabification, the options for Castilian prosody have been enabled, as well as epentheses. The analysis takes 87 seconds (±1.5 s) in all cases (running Stanza on the GPU), disregarding the approximately 7 seconds Stanza takes to load.

In short, libEscansión scores 97.01% accuracy against the corpus when running with its standard settings (Castilian syllabification), an improvement over previous scansion engines (Table 3). The American syllabification renders slightly worse results, with 96.51%. However, as noted above, this corpus considers interjections such as oh and ay as unstressed, whilst our library does not. This behaviour can be reproduced with an ad hoc three-line hack in the code, in which case the accuracy would reach 98.65% for the standard setting and 98.15% for the American syllabification. We have considered inconsistencies in the corpus notation for the test as some proparoxytonic verses of the corpus are marked by omitting one of the last two unstressed syllables.

Apart from the disagreements due to interjections, there are 19 disparate verses left for other reasons and another with disagreements beyond the interjections (Table 4, row 18). There are ten disagreements due to spelling deviations (rows 1–5) and annotator mistakes (rows 6–12). In the first case, the manual annotation reflects the pronunciation despite the lack of accents, such as in words like Príamo and the more subtle diacritics of aún and dó. Consequently, humans have an advantage over machines as the former can recognise the proper pronunciation despite the spelling. The second case ranges from non-marked stresses in content words, disregarding the primary stress of an adverb in -mente or an adjective, or the opposite, stressing an unstressed syllable like in an honorary title preceding a noun or the locution ya que.

Two disagreements (rows 18–19) deserve a special mention. The first is due to a PoS-tagging mistake, as para has been tagged as a preposition due to the lack of a vocative comma, which illustrates what we said above. Even though other potential errors of a similar kind have been intercepted, ambiguities resembling row 18 cannot be bypassed since the PoS-tagger has operated as expected, and it is unviable to predict any potential typo or edition mistake. The second mistake (row 19) is rather interesting. The syllabificator breaks the diphthong, but VerseMetre, after pondering all the options, joins the vowels again as it prioritises placing the caesura between words when keeping editorial diareses is not explictly forced.

If we leave the remaining inter-annotator disagreements, be they errors or different yet valid takes on the verse, libEscansión reaches a 99.50% accuracy.