Normalization of Non-Standard Words with Finite State Transducers
for Russian Speech Synthesis

Text Preprocessing for Speech Synthesis

• is usually a very complex task
• Text normalization is one of the steps in text preprocessing [1]
  • sentence segmentation
  • tokenization
  • normalization of non-standard words (NSWs)
    • numbers, abbreviations, and acronyms
    • different characters like %, $, #, №, etc.

Normalization of Non-Standard Words

• NSWs must be expanded into full SW to be pronounced correctly
• It's even more complex in inflective languages such as Russian
  • ordinal number can be converted into 36 different word forms (6 cases * 2 numers * 3 genders)
  • digit position changes the output standard word
    • 11111 — первый
    • 11111 — одиннадцатый
    • 11111 — сто
    • 11111 — тысяча
    • 11111 — одиннадцать тысяч

Existing Russian Normalization Systems

• As a part of proprietory Text-to-Speech (TTS) systems
  • Google Translate, https://translate.google.ru/
  • VitalVoice, http://cards.voicefabric.ru/
  • Windows SAPI voices, etc.
• As a part of open-source TTS systems
  • Festival [2]
    • only digit-by-digit number normalization for the Russian voice

Normatex

• is the first Russian open-source normalization system, known to the author, github.com/avlukanin/normatex
• If the input texts are normalized beforehand the quality of the synthesized speech of existing TTS systems can be improved
• 118 finite state transducers (FSTs) for conversion of cardinal and ordinal numbers into the corresponding numerals, which can preprocess different ranges, time, dates, telephone numbers, postal codes, etc.
• 33 FSTs for normalization of graphic abbreviations and acronyms

Test Parallel Corpus

• 66 original texts of the official site of South Ural State University, susu.ac.ru, which contains 38,439 tokens (broad segmentation units [3]):
  • 14,661 word tokens
    • 333 acronyms and 98 initials; 379 graphic abbreviations
  • 977 number tokens (2,511 digits)
• 66 manually preprocessed texts, where all numbers, abbreviations and acronyms were expanded into full words or replaced with pronounceable combination of letters

Finite State Transducers

• are developed in the form of graphs in Unitex 3.1beta
• Before applying FSTs to a text, it is preprocessed:
  • The text is splitted into sentences
  • The text is tokenized
  • Every token is assigned all possible grammatical forms
• Number FSTs are applied first to deal with numbers and measure unit abbreviations
• Abbreviation FSTs and acronym FSTs are applied sequentially after that

Cardinal Numbers

• agree with nouns in case, but the numerals один “one” and два “two” agree in gender as well
• all the constituent words of a compound numeral agree with the corresponding noun: двадцати одного and двадцати одной (“twenty-one” in gen. m. and f.)
• одни (“one” in plural) agrees only with pluralia tantum, e.g. одни ножницы “one pair of scissors”, одни брюки “one pair of pants” [4]

5-9ncard

2x-9xncard

NUM-5-9-ncard

units

Ordinal Numbers

• Simple ordinal numerals agree with nouns in gender, case and number
• In compound ordinal numerals only the last constituent word agrees with the noun [5]: две тысячи четырнадцатом (“two thousand fourteenth” in prepositional masculine)
• Complex ordinal numbers, ending in -00, -000, -000000, -000000000, are written without spaces: “153000” is converted into стопятидесятитрёхтысячный “one hundred and fifty-three thousandth” in nominative masculine

Ordinal Numbers

• Only the last constituent words -сотый “hundredth”, -тысячный “thousandth”, -миллионный “millionth”, -миллиардный “billionth” agree with the nouns
• The words, preceding the last word, are used in genitive plural (the exceptions are сто “one hundred” and девяносто “ninety”, which are used in the nominative case) [6]

Acronyms

• Most acronyms should be converted into full words before speech synthesis, because it is difficult for people to comprehend a letter-by-letter pronunciation in speech and because acronyms are often rare for everybody to know what phrase the acronym corresponds to

ФГБОУ ВПО «ЮУрГУ» (НИУ)

ФГБОУ ВПО «ЮУрГУ» (НИУ) → Федеральное государственное бюджетное образовательное учреждение высшего профессионального образования «Южно-Уральский государственный университет» (Научно-исследовательский университет)

Acronyms

• The main component of an acronym is a noun, that is why there can be 12 possible forms of the converted phrase (six cases and two numbers) in Russian
• There are rules for all six cases in Normatex
• Acronyms can be ambiguous in different corpora
• For all ambiguous or unknown acronyms Normatex substitutes each letter with its alphabet name: ВПП → ВэПэПэ

Graphic Abbreviations

• Single interpretation: и т.д. “etc.” → и так далее, т.е. “i.e.” → то есть
• The interpretation depends on the context: и др. “et al.” → и другие “and others”, и других “and others”, и другим “and others”, и другое “and other”
• Ambiguous: г. → год “year”, город “city”, грамм “gram” (every noun can have 12 word forms), Аудитория: 339-г, 339-д “Room 339-g, 339-d
• Sufficient left and right contexts should be provided in FSTs as well as FSTs should be applied in a definite order

Results

Token type	Tokens	Correct	Errors	Recall	Precision
Numbers	977	920	53	94.17%	94.55%
Acronyms and initials	431	355	40	82.37%	89.87%
Graphic abbreviations	379	232	4	61.21%	98.05%
Total	1787	1507	97	84.33%	93.95%

The work is still in progress

References

1. Reichel, U.D., Pfitzinger, H.R.: Text preprocessing for speech synthesis (2006)
2. The Festival Speech Synthesis System, http://www.cstr.ed.ac.uk/projects/festival/
3. Dutoit, T.: An introduction to text-to-speech synthesis (Vol. 3). Springer Science & Busi-ness Media (1997)
4. Russian Grammar [Русская грамматика]. Vol. 1. Nauka, Moscow (1980)

References

5. Rosental, D.E., Golub, I.B., Telenkova, M.A.: The Modern Russian Language [Современный русский язык]. Airis-Press, Moscow (1997)
6. Rosental, D.E., Djandjakova, E.V., Kabanova, N.P.: Reference Book on Orthography, Pronunciation, Literary Editing [Справочник по правописанию, произношению, литературному редактированию]. CheRo, Moscow (1998)

Normatex — Russian text normalization

github.com/avlukanin/normatex

Artem Lukanin

• about.me/alukanin
• @avlukanin
• artyom.lukanin@gmail.com

Slides: artyom.ice-lc.com/slides/normatex