MFA IPA phone set#
For its IPA models, MFA uses an opinionated IPA phone set. This page will catalog the way that phones are represented across languages (and note any deviations from them). The guiding principle is to have general consistency in phones across languages, but allow for phonetic specification where it A) matters for the language and B) aids the aligner in assigning HMM states to phones.
For most languages, dictionaries were constructed from the Wikipron scraped dictionaries. However, given that they are crowd-sourced transcriptions from Wiktionary, there was a fair bit of noise to clean up.
Warning
This page is under heavy construction as more languages get trained, so many languages are missing or have incomplete entries below.
General notes#
Diacritic ordering: The ordering of diacritics is (prenasalization) symbol (secondary place) (glottal settings) (length)
pʰː
pʷʼ
pʲ̚
ⁿd̪ʷʱː
Removal of narrow transcription diacritics:
Height markers on vowels and consonants
[o̞] [o] in various languages
[ɾ̝̊] [ɾ] in Turkish
[ɯ̟̃ᵝ ɨ̥ᵝ] [ɯ] or [ɨ] depending on surrounding context in Japanese
Removal of nasality in languages where it’s not phonemic
Merging some non-phonemic or dialectal contrasts
[ʍ] [w] in English
[r] [ʁ] in French
[ʁ ɹ ɻ χ ɦ h r] [x] in Brazilian Portuguese
Standardizing diphthongs, afficates, etc
Merging sequences of [a ʊ] into a single [aw] phone
Merging sequences of vowels in Mandarin to single nucleus
Standardizing tones
Transforming all tones to Chao system (i.e. Mandarin: [k w a²¹⁴ n] [k w a˨˩˦ n])
Placing tone diacritics on the vowel nucleus rather than on the final segment in the syllable
Fixing up character glyphs
[õ ẽ ũ ĩ ã] [õ ẽ ũ ĩ ɐ̃] in Portuguese (single glyph to compound characters with nasality diacritic)
Correcting many mistranscriptions
[r] [ɹ] in US English
[cʰ] [tɕʰ] in Thai
[tʃ] [tɕ] in Tamil
[ç] [ɕ] in Swedish
I have avoided having numeric characters in the pronunciation dictionaries, generally converted to the language’s orthography via the hanziconv package.
Note
I am unfamiliar with many of the languages and dialects below, and have largely based my systems on publicly available resources, relying heavily on any phonology information on Wikipedia, as well as Phoible and XPF. If you have any suggestions for improvements, please open a discussion topic.
Consonants#
Palatalized consonants: If a language has palatalized consonants, these have been replaced by palatal consonants:
[lʲ] [ʎ]
[nʲ] [ɲ]
[kʲ] [c]
[ɡʲ] [ɟ]
[xʲ] [ç]
[ɣʲ] [ʝ]
Place representation: In general, if a consonant is listed as dental in descriptions of its typical realization, it is represented with the dental diacritic
[t̪ d̪ t̪s̪ n̪]
Homorganic nasal place assimilation: If a language has this (and let’s be real, which ones don’t?[1]), then the following transformations are done:
[m] [ɱ] before labiodental consonants
[n ŋ] [ɲ] before palatal consonants
Vowels#
Diphthongs with high on-glides or off-glide have been standardized to use glides rather than full vowels or marked with brevis diacritics
[au aʊ aʊ̯] [aw]
[ou oʊ oʊ̯] [ow]
[ai aɪ aɪ̯] [aj]
[ei eɪ eɪ̯] [ej]
Specific language notes#
Arabic#
The Arabic MFA dictionary was created by Natalia Shmueli.
Pronunciation dictionaries
See Arabic MFA dictionary v2.0.0 for full IPA charts.
Arabic consonants#
Gemination: Multiple or geminated consonants are represented as long consonants
[l l] [ɫː]
Arabic vowels#
Short vowel laxing: Short vowels are represented with their lax variants, while long vowels retain their tense variants
[a] [æ], except before [q sˁ tˁ dˁ ðˁ ɫ r] where it is realized as [ɑ]
Long [a] is [aː], except before [q sˁ tˁ dˁ ðˁ ɫ r] where it is realized as [ɑː]
[i] [ɪ], but not for [iː]
[u] [ʊ], but not for [uː]
Bulgarian#
Pronunciation dictionaries
See Bulgarian MFA dictionary v2.0.0 for full IPA charts.
Bulgarian consonants#
Dental consonants: Alveolar obstruents and nasals are represented with dentals [t̪ d̪ t̪s̪ s̪ z̪ n̪], except when palatalized
Bulgarian vowels#
Unchanged from their representation in Wikipron.
English#
The primary deviations from standard symbols have been to bring English more in line with other languages.
Pronunciation dictionaries
See English (India) MFA dictionary v3.0.0, English (Nigeria) MFA dictionary v3.0.0, English (UK) MFA dictionary v3.0.0, and English (US) MFA dictionary v3.0.0 for full IPA charts.
English consonants#
Palatalization: Changed [n k ɡ l h] tokens to [ɲ c ʎ ɟ ç] before front vowels [i(ː) ɪ ej ɛ æ], as well as before other palatal consonants
Yod coalescence: Combined some sequences involving [j] to a single palatal symbol
[n j] [ɲ]
[k j] [c]
[l j] [ʎ]
[ɡ j] [ɟ]
[h j] [ç]
Aspiration: Added aspiraton to voiceless stops before stressed vowels and at the beginning of words, except after sibilants
[p] [pʰ]
[t] [tʰ]
[c] [cʰ]
[k] [kʰ]
Flapping (US only): Changed [t d] tokens to [ɾ] in unstressed syllables, except after nasals
rapidity [ɹ ə pʰ ɪ ɾ ɪ ɾ i]
Intervocalic /t/ glottalization (UK only): Added pronunciation variants for [t] [ʔ] in unstressed syllables, except after nasals and obstruents
butter [b ɐ t ə] ~ [b ɐ ʔ ə]
Word-final /t/ glottalization (all dialects): Added pronunciation variants for word-final [ʔ] after vowels and liquids
right [ɹ aj t] ~ [ɹ aj ʔ]
Initial glottals: Added pronunciation variants with initial [ʔ] for common vowel-initial words
Dental stopping (all dialects): Added pronunciation variants for realizations of [ð θ] as [d t] for frequent words
English vowels#
goose-fronting (US and UK): Changed [u] tokens to the more fronted [ʉ], Nigerian English retains [u]
dude [dʲ ʉː d]
lot, cloth, thought-lowering (US and UK): Changed [ɔ] tokens to the lower [ɒ]
caught [kʰ ɒː t]
strut-centering (US and UK): Changed [ʌ] tokens to the more central [ɐ], Nigerian English has these tokens as [ɔ]
but [b ɐ t]
merry/marry/mary merger (US only): [ej æ ɛ] are merged to [ɛ] before [ɹ]
Diphthong standardization (all dialects): Changed diphthongs to use glides rather than extra short lax vowels
[aʊ] [aw]
[aɪ] [aj]
[oʊ] [ow]
[eɪ] [ej]
[ɔɪ] [ɔj]
Unstressed vowels (US and UK): I have tried to limit the variation of [ɪ] and [ə] in unstressed syllables by confining it to [ɪ] in the -ed (verb past tense) and -es (plurals and third person present tense). Elsewhere it’s harder to limit, but if two variants only differ in quality between [ɪ] and [ə], I’ve only kept one of them
Diphthong + rhotic standardization: In most words that have a diphthong plus rhotic, these have been standardized to use [ɹ] rather than [ɚ], except when there is an -er affix
fire [f aj ɚ] fire [f aj ɹ]
But higher [h aj ɹ] higher [h aj ɚ]
Nigerian English#
The Nigerian English was based on the UK dictionary, with some significant modifications. I have tried to largely follow Ulrike B. Gut’s book chapter on Nigerian English phonology[2], as that is what is predominately cited in other resources like the OED.
Pronunciation dictionaries
See English (Nigeria) MFA dictionary v3.0.0 for full IPA charts.
Nigerian English consonants#
Silent consonants: Silent consonants in words like bomb, numb, etc are realized, i.e. [b ɔ m b]
Realization of unstressed [t]: Realized as [t] rather than [ʔ] in the UK dictionary
Borrowed labial-velar stops (As of 3.0.0)
ogbomosho [ɔ ɡb o m o ʃ o]
ukpabi [ʊ kp a bʲ i]
Nigerian English vowels#
[ɔ]: lot, strut, cloth, north, force, thought
lot [l ɔ t]
strut [s t ɹ ɔ t]
cloth [k l ɔ θ]
north [n ɔ θ]
force [f ɔ s]
thought [θ ɔ t]
[a]: trap, bath, palm, start, letter (unstressed rhotic), unstressed [ə] for orthographic a
trap [t ɹ a p]
bath [b a θ]
palm [pʰ a m]
start [s t a t]
letter [l ɛ t a]
about [a b aw t]
[i]: kit, fleece, unstressed [ə] or [ɪ] for orthographic i
kit [cʰ i t]
fleece [f ʎ iː s]
singing [s i ŋ i ŋ ɡ]
[u]: goose
goose [ɡ uː s]
[e]: face
face [f e s]
[o]: goat
goat [ɡ o t]
[ɛ]: dress, unstressed [ə] or [ɪ] for orthographic e, especially -ed and -es suffixes
dresses [d ɹ ɛ s ɛ z]
granted [ɡ ɹ a n t ɛ d]
[ʊ]: foot, cure, unstressed syllables with syllabic [ɫ̩], especially -ful and -able suffixes
foot [f ʊ t]
cure [cʰ ʊ a]
wonderful [w ɔ n d a f ʊ ɫ]
legible [l ɛ dʒ i b ʊ ɫ]
Indian English#
The Indian English dictionary is largely based on the UK dictionary with modifications following the Wikipedia page on Indian English phonology.
Pronunciation dictionaries
See English (India) MFA dictionary v3.0.0 for full IPA charts.
Indian English consonants#
Retroflex alveolar stops:
[t] [ʈ]
[d] [ɖ]
Labiodental realization of [w]
whale [w eː l] [ʋ eː l]
Stopped realization of dental fricatives
other [ə ð ə] [ə d̪ ə]
think [θ ɪ ŋ k] [t̪ ɪ ŋ k]
Indian English vowels#
face: [eː]
goat: [oː]
trap: [a]
strut: [ə]
Added pronunciation variants for flapping of final rhotic schwa
never [n ɛ ʋ ə] [n ɛ ʋ ə ɾ]
English phone groups#
Phone group configuration
See English phone group configruation for exact specification.
English phonological rules#
Phonological rule configuration
See English phonological rule configruation for exact specification.
Hausa#
Largely followed the Hausa phonology wiki.
Pronunciation dictionaries
See Hausa MFA dictionary v3.0.0 for full IPA charts.
Hausa vowels#
Tone: Changed tone makers to use tone level symbols following vowels rather than accent diacritics
abacada [ʔ àː b àː t͡ʃ àː d âː] abacada [ʔ aː˩ b aː˩ tʃ aː˩ d aː˥˦]
Diphthongs: Merged instances of [i] and [u] without tone following [a] and [e] into diphthongs [aj] and [aw]
Japanese#
Tokenization
The Japanese MFA acoustic model v3.0.0 tokenizes transcripts automatically using sudachipy rather than relying on pre-tokenized transcripts. The changes below are implemented in montreal_forced_aligner.tokenization.japanese.
Japanese does not represent word breaks in text, and is a heavily agglutinative language, so a pronunciation dictionary with discrete pronunciations isn’t the best model for it. As such for the purposes of building a lexicon and acoustic model, I used the Nagisa morphological analyzer to split text for all corpora, even ones where there was some word break information to be consistent. However some postprocessing of the nagisa-split transcripts was necessary:
If a “word” ended in っ (chisai-tsu), then the following “word” was appended
行っ て 行って
Realization of っ is generally dependent on the following consonant
If a つ character was by itself and the previous word was a numeral 1-10, then the two words were combined
一 つ 一つ
Native counters with つ result in wholesale changes to the pronunciation, e.g. 一 [i tɕ i] 一つ [ç i̥ t o ts ɨ]
Any roman numerals were converted to Kanji
2010年 二千二十年
Kanji pronunciations try to include all onyomi and kunyomi variants in case there’s an issue with the morphological parse
Pronunciation dictionaries
See Japanese MFA dictionary v3.0.0 for full IPA charts.
Japanese consonants#
Gemination: Geminate consonants are uniformly represented with length marking
[tː ɸː pʲː mʲː]
Palatalized consonants: Palatalized velar consonants, as well as those before * [i] were coded as fully palatal
がっきゅう [ɡ a cː ɨː]
ちょうき [tɕ oː c i]
きぎょう [c i ɟ oː]
ぎゅうにゅう [ɟ ɨː ɲ ɨː]
Voiced fricative stopping: [z] and [ʑ] are affricates [dz] and [dʑ] word-initially, following nasals, or when geminated
じょうでき [dʑ oː d e c i]
マッジ [m a dʑː i]
あんぜん [a n dz e ɴ]
Realizations of ん: Coda nasal is realized as:
[m] before [p b] (and their palatalized/geminate variants)
[n] before [t ts dz d ɾ] (and their palatalized/geminate variants)
[ɰ̃] before [s ɕ ɸ ç h] (and their palatalized/geminate variants) and vowels
[ɲ] before [c ɟ tɕ dʑ ɕ ʑ] (and their geminate variants)
[ŋ] before [k ɡ] (and their geminate variants)
[ɴ] word-finally
Long nasal before other nasals, i.e. うんめい [ɯ mː eː]
Realizations of /h/:
[ɸ] before [ɯ] (and its long variant)
[ç] before [i] (and its long variant) and in palalized cases like ひゃく [ç a k ɯ]
[h] elsewhere
Borrowed sounds: The kana ヴ is transcribed with [v]
アヴェニュー [a v e ɲ ɨː]
Final glottals: Words ending in っ that are pronounced with audible stops are transcribed with a final [ʔ]
あっ [a ʔ]
いてっ [i t e ʔ]
Japanese vowels#
[ɯ]-fronting: Following alveolar obstruents or palatal(ized) consonants, [ɯ] [ɨ]
High vowel devoicing: Between voiceless obstruents, the high short vowels [i ɨ ɯ] are represented as voiceless [i̥ ɨ̥ ɯ̥]
すき [s ɨ̥ c i]
High frequency words like です, ます, etc. have a pronunciation variant with devoiced final vowel です [d e s ɨ̥] ~ [d e s ɨ]
Japanese phone groups#
Phone group configuration
See Japanese phone group configruation for exact specification.
Japanese phonological rules#
Phonological rule configuration
See Japanese phonological rule configruation for exact specification.
High vowel devoicing between voiceless obstruents
好き [s ɨ c i] [s ɨ̥ c i]
しかし [ɕ i k a ɕ i] [ɕ i̥ k a ɕ i]
ふたつ [ɸ ɯ t a ts ɨ] [ɸ ɯ̥ t a ts ɨ]
High vowel deletion between voiceless obstruents
好き [s ɨ c i] [s c i]
しかし [ɕ i k a ɕ i] [ɕ k a ɕ i]
ふたつ [ɸ ɯ t a ts ɨ] [ɸ t a ts ɨ]
High vowel devoicing word-finally
しかし [ɕ k a ɕ i] [ɕ k a ɕ i̥]
です [d e s ɨ] [d e s ɨ̥]
とにかく [t o ɲ i k a k ɯ] [t o ɲ i k a k ɯ̥]
High vowel deletion word-finally
しかし [ɕ k a ɕ i] [ɕ k a ɕ]
です [d e s ɨ] [d e s]
とにかく [t o ɲ i k a k ɯ] [t o ɲ i k a k]
Korean#
Tokenization
The Korean MFA acoustic model v3.0.0 tokenizes transcripts automatically using python-mecab-ko rather than relying on pre-tokenized transcripts. There are some additional rules implemented in montreal_forced_aligner.tokenization.korean.
I have largely followed the Korean Phonology Wikipedia page. Given the agglutinative nature of Korean, I have standardized text in corpora by processing them through the Mecab-KO morphological parser using the KoNLPy package. Additionally, I have included G2P pronunciations of the phonological transcriptions in :need:Seoul Corpus.
Pronunciation dictionaries
See Korean MFA dictionary v3.0.0 for full IPA charts.
Korean consonants#
Palatalization:
ㅅ [s sʰ] [ɕ ɕʰ] and ㅆ [s͈] [ɕ ɕʰ] [ɕ͈ ] before [i j]
Diphthongs with onglides (ㅖ ㅒ ㅑ ㅛ ㅠ ㅕ ㅟ ㅞ ㅙ ㅘ ㅝ) are realized as palatalization or labialization on the preceding segment rather than separate glide (As of version 3.0.0)
려 [ɾ j ʌ] [ɾʲ ʌ]
쇠사슬 [sʰ w e sʰ ɐ sʰ ɨ ɭ] [sʷ e sʰ ɐ sʰ ɨ ɭ]
Segments preceding [i j] are realized as palatalized (As of version 3.0.0)
기업 [k i ʌ p̚] [c i ʌ p̚]
미터 [m i tʰ ʌ] [mʲ i tʰ ʌ]
Gemination
Sequences of unreleased stops following by same place obstruents are realized as long consonants (As of version 3.0.0)
체육관 [tɕʰ e j u k̚ kʷ ɐ n] [tɕʰ e j u kʷː ɐ n]
Realizations of ㅎ /h/:
[ç] word-initially before [i j]
[ʝ] intervocalically before [i j]
[x] word-initially before [ɯ]
[ɣ] intervocalically before [ɯ]
[ɸ] word-initially before [o u w]
[β] intervocalically before [o u w]
[h] word-initially before [o u w]
[ɦ] intervocalically before [ɐ ʌ e]
Realizations of ㄹ /l/:
[ɾ] word-initially, intervocalically, or between vowels and /h/
[ʎ] before [i j] or palatal consonants
[ɭ] elsewhere
Korean vowels#
[a]-raising: Open vowels are transcribed with as [ɐ]
/ɯ/-fronting: /ɯ/ is transcribed as /ɨ/
Vowel length: Vowel length marking is retained as sourced from Wikipron; however, it seems like it’s completely neutralized in modern Korean
Diphthongs: Diphthongs are transcribed as sequences of independent glide + vowel combinations
가격의 [k ɐ ɡ j ʌ ɡ ɰ i]
가는쥐 [k ɐ n ɨ ɲ dʑ ɥ i]
가두녀성 [k ɐ d u ɲ j ʌ sʰ ʌ ŋ]
가시화될 [k ɐ ɕʰ i β w ɐ d w e ɭ]
Korean Phone groups#
Phone group configuration
See Korean phone group configruation for exact specification.
Korean phonological rules#
Phonological rule configuration
See Korean phonological rule configruation for exact specification.
Deletion of unreleased stops word-finally
기획 [c i βʷ e k̚] [c i βʷ e]
Deletion of unreleased stops before obstruents with other places
낙찰 [n ɐ k̚ tɕʰ ɐ ɭ] [n ɐ tɕʰ ɐ ɭ]
Deletion of /h/ after /n/
뻔히 [p͈ ʌ n ʝ i] [p͈ ʌ ɲ i]
분할 [p u n h ɐ ɭ] [p u n ɐ ɭ]
Denasalization of m n word-initially
낙찰 [n ɐ k̚ tɕʰ ɐ ɭ] [d ɐ k̚ tɕʰ ɐ ɭ]]
밑 [mʲ i t̚] [bʲ i t̚]
Affrication of palatalized alveolar stops
버팀목 [p ʌ tʲ i mː o k̚] [p ʌ tɕʰ i mː o k̚]
부디 [p u dʲ i] [p u dʑ i]
Mandarin#
Tokenization
The Mandarin MFA acoustic model v3.0.0 tokenizes transcripts automatically using spacy-pkuseg rather than relying on pre-tokenized transcripts. There are some additional rules implemented in montreal_forced_aligner.tokenization.chinese.
Orthographic text transcriptions for Mandarin were segmented with the spacy-pkuseg package. Orthographic texts for Mainland Chinese corpora, Mandarin (China) MFA dictionary v2.0.0, and Mandarin (Erhua) MFA dictionary v2.0.0 use simplified characters, while corpora for Taiwanese Mandarin and Mandarin (Taiwan) MFA dictionary v2.0.0 use traditional characters (converted where necessary with hanziconv). Roman numerals were converted to Chinese characters via num2chinese.py. As of 3.0, erhua pronunciations have been combined into Mandarin (China) MFA dictionary v3.0.0.
Building the Mandarin dictionaries followed a slightly different process than the other dictionaries. The base Wikipron dictionary contains forms collapsed across various dialects of Mandarin, as “Mandarin” itself functions as the dialect of Chinese on Wiktionary. I wrote a custom scraping script that funneled pronunciations into three dictionaries. Prior to 3.0 models, if a pronunciation was marked with “Taiwan” or “Erhua”, then it is present only in those respective dictionaries. In 3.0 models, erhua pronunciations are included in Mandarin (China) MFA dictionary v3.0.0 without a separate dictionary. Any pronunciation marked with just “Standard Chinese” was added to all dictionaries.
For generating new pronunciations, the standard G2P process used for other languages does not work well (single characters correspond to multiple phones in a non-decomposable way). To get around this, I wrote a custom script that assigns syllables to characters and compiles frequency representations of each character and all sequences of characters. Pronunciations generated for a novel compound is the combination of the longest matched subsequences. There are likely errors from tone sandhi or other phonological effects, but they should be represented by at least one combination of pronunciations, and by reusing common longer sequences, but it’s definitely not perfect.
Pronunciation dictionaries
See Mandarin (China) MFA dictionary v3.0.0,and Mandarin (Taiwan) MFA dictionary v3.0.0 for full IPA charts.
Mandarin consonants#
Mandarin vowels#
Tone variation: In general, I have left whatever variation was present in Wikipron.
Diphthongs: As with other languages, diphthongs with high vowels are represented as having off-glides [ej ow aw aj].Initial glids are separate from the vowel nucleus.
㑿 [ʈʂ aw˥˩]
㒟 [n j aw˨˩˦]
㑼 [l ɥ e˥˩]
Mandarin Phone groups#
Phone group configuration
See Mandarin phone group configruation for exact specification.
Mandarin phonological rules#
Phonological rule configuration
See Mandarin phonological rule configruation for exact specification.
Deletion of glottal stops
[] []
Deletion of erhua ɻ
[] []
Polish#
Pronunciation dictionaries
See Polish MFA dictionary v2.0.0 for full IPA charts.
Polish consonants#
Dental consonants: Coronal consonants are represented with dental diacritics, except when palatalized
[t̪ tʲ d̪ dʲ t̪s̪ tsʲ n̪ ɲ]
Nasal approximation: Palatal nasal [ɲ] becomes [j̃] before fricatives
Polish vowels#
Portuguese#
Pronunciation dictionaries
See Portuguese (Brazil) MFA dictionary v2.0.0 and Portuguese (Portugal) MFA dictionary v2.0.0 for full IPA charts.
Portuguese consonants#
Palatal nasal approximation (Brazilian Portuguese only): Converted all nh [ɲ] to [j̃]
Palatalization before high vowels/glides:
[l] [ʎ] and [n] [ɲ] before [i] and [ĩ]
[l j] [ʎ] and [n j] [ɲ]
[d] [dʒ] and [t] [tʃ] before front vowels
Portuguese vowels#
Diphthongs are treated as sequences of vowel + glide, i.e., limão [ʎ i m ɐ̃ w̃] (Brazilian)
Rationale is that there is a high degree of combinations between vowels and glides, with some combinations only represented a handful of times in the lexicons
Russian#
Pronunciation dictionaries
See Russian MFA dictionary v2.0.0 for full IPA charts.
Russian consonants#
Dental consonants: Hard coronal consonants are represented with dental diacritics, but not soft coronal cosonants
[t̪ tʲ d̪ dʲ t̪s̪ tsʲ n̪ ɲ]
Russian vowels#
Unedited from Wikipron.
Spanish#
Pronunciation dictionaries
See Spanish (Latin America) MFA dictionary v2.0.0 and Spanish (Spain) MFA dictionary v2.0.0 for full IPA charts.
Spanish consonants#
Voiced stop lenition (all dialects): Voiced stops are represented as fricatives in all places except word-initially, after a nasal consonant, and after a lateral consonant (for /d ʝ/ only)
[#n] [b d̪ ɡ ɟʝ] [^#n] [β ð ɣ ʝ]
Palatalization (all dialects): Following other languages, velars [k ɡ x ɣ] [c ɟ ç ʝ] before [i e]
Dental obstruents (all dialects): Coronal stops are represented with a dental diacritic (other manners for coronals are represented as alveolar)
[t̪ d̪ s n]
Yeísmo: ll [ʎ] and y [ʝ] are merged in the Latin American dictionary, but remain split in the European one. However, it might be worth merging them for European as well as that is a sound change in progress.
Spanish vowels#
Unedited from Wikipron.
Swahili#
Pronunciation dictionaries
See Swahili MFA dictionary v2.0.0 for full IPA charts.
Swahili consonants#
Prenasalized obstruents: Treated as sequences of nasal + obstruent
ndewe [n d ɛ w ɛ]
Implosives: Voiced stops that are not following nasals are represented as implosives
mbabe [m b ɑ ɓ ɛ]
Aspiration: Voiceless stops are all unaspirated, as it is not reflected in the orthography and I do not have a definitive source for which words contain aspirated stops vs plain
Trills: All instances of /r/ are represented with a tap [ɾ]
Swahili vowels#
Vowel length: Sequences of vowels are treated as long vowels
Swedish#
Pronunciation dictionaries
See Swedish MFA dictionary v2.0.0 for full IPA charts.
Swedish consonants#
Dental consonants: Alveolar obstruents and nasals are represented as dental [t̪ d̪ s̪ n̪]
Swedish vowels#
Pitch accent: Pitch accent markings for words on Wiktionary (superscript 1 or 2) and pitch accent markings on vowels have been expanded to use IPA tones
gifter /¹ j ɪ f t ɛ r/ [j ɪ˥˧ f t̪ ɛ˩ r]
gifter /² j ɪ f t ɛ r/ [j ɪ˧˩ f t̪ ɛ˥˩ r]
ära [ɛ̂ː r a] [ɛː˧˩ r a˥˩]
Thai#
Tokenization
The Thai MFA acoustic model v3.0.0 tokenizes transcripts automatically using pythainlp rather than relying on pre-tokenized transcripts. There are some additional rules implemented in montreal_forced_aligner.tokenization.thai.
Pronunciation dictionaries
See Thai MFA dictionary v3.0.0 for full IPA charts.
Thai consonants#
Unreleased final stops: Coda stops are transcribed as unreleased [p̚ t̚ k̚]
คร้าบ [kʰ r aː˥˩ p̚]
Thai vowels#
Tones: Tones have been attached to the vowel of the syllable
คลอก [kʰ l ɔː˥˩ k̚]
Diphthongs: Glides have been analyzed as separate segments
กบไสไม้ [k o˨˩ p̚ s a˩˩˦ j m aː˦˥ j]
จุดไข่ปลา [tɕ u˨˩ t̚ kʰ a˨˩ j p l aː˧]
Thai phone groups#
Phone group configuration
See Thai phone group configruation for exact specification.
Thai phonological rules#
Phonological rule configuration
See Thai phonological rule configruation for exact specification.
Deletion of unreleased stops
เกรด [k r eː˨˩ t̚] [k r eː˨˩]
Deletion of glottal stops
เกาะ [k ɔ˨˩ ʔ] [k ɔ˨˩]
Turkish#
Largely followed the Turkish phonology Wikipedia page, following their list of phonetic realizations.
Pronunciation dictionaries
See Turkish MFA dictionary v2.0.0 for full IPA charts.
Turkish consonants#
Palatalization: Velar obstruents are palatal [c ɟ ç] before front vowels [i e œ y], as well as the alveolar lateral [ʎ]
Homorganic nasal place assimilation: Nasals agree in place with following obstruents, transcribed as one of [m n̪ ɲ ŋ]
Dental consonants: Alveolar stops and nasals are transcribed as dental [t̪ d̪ s̪ z̪ n̪]
Gemination: Sequences of two consonants in the orthography are represented as a long consonant
minnettar [m i n̪ː e t̪ː a ɾ]
Geminated [ɾ] is represented as [ɾː], but maybe should be trilled [r]
Realizations of ğ /ɰ/: Given its quasi-phonemic status, I’ve followed the description of its phonetic distribution:
Before a consonant or word boundary, it is realized as length on the previous vowel
Between front vowels, it is realized as [j]
Words with ğ /ɰ/ between back vowels have two pronunciation variants, one with [ɰ] realized and one with it deleted sağol [s̪ a ɰ o ɫ] ~ [s̪ a o ɫ]
Turkish vowels#
Phonetic realizations: The primary realization of front vowels is [i y e ø], but for word-final open syllables, these are represented with [ɪ ʏ ɛ œ]. Back vowels are generally represented as [ɯ u a o], with word-final open realizations for the high back vowels are [ɨ ʊ]
Closed syllable [e] lowering: In syllables with coda [m ʎ ɲ ɾ ŋ ɫ n̪], [e] [ɛ]
Ukrainian#
Pronunciation dictionaries
See Ukrainian MFA dictionary v2.0.0 for full IPA charts.
Ukrainian consonants#
Palatalization: Velar and alveolar soft consonants are represented as palatal
[kʲ] [c]
[ɦʲ] [ç]
[xʲ] [ʝ]
[nʲ] [ɲ]
[lʲ] [ʎ]
Trill realization: The trill /ʝ/ is represented as [ɾ]
Dental consonants: Alveolar consonants are dental [t̪ s̪ t̪s̪ d̪ z̪ d̪z̪ n̪] except when palatalized
Ukrainian vowels#
Largely unchanged from their Wikipron counterparts.
Vietnamese#
Generally I have followed James Kirby’s vPhon for generating transcriptions of Vietnamese words across three dialects. The points where I have diverged from Kirby’s system are below.
Pronunciation dictionaries
See Vietnamese (Hanoi) MFA dictionary v2.0.0, Vietnamese (Ho Chi Minh City) MFA dictionary v2.0.0, and Vietnamese (Hue) MFA dictionary v2.0.0 for full IPA charts.
Vietnamese consonants#
Palatalization: Converted palatalized segments in Hanoi dialect to palatal consonants
đánh [ɗ a˨˦ ʲŋ] [ɗ a˨˦ ɲ]
thạch [tʰ a˨˩ ʲk] [tʰ a˨˩ c]
Changing final consonants to be unreleased
bạc [ɓ aː˨˩ k] [ɓ aː˨˩ k̚]
IPA cleanup: Removed ligatures and normalized bilabial onglides
hóc [h ɔ˦˥ k͡p] [h ɔ˦˥ kp]
đông [ɗ o˨˨ ŋ͡m] [ɗ o˨˨ ŋm]
hoan [h ʷ aː˨˨ n] [h w aː˨˨ n]
Vietnamese vowels#
Tones: Converted Chao numbers into IPA levels and placed them always on the vowel rather than at the end of the syllable
chải [tɕ aː j ²¹²] [tɕ aː˨˩˨ j]
[^Rikker] I appreciate Rikker Dockum (website) taking the time to answer my questions about various aspects of Thai and taking a look over some initial versions of the pronunciation dictionary! Errors are still my own.