Error analysis and improving speech recognition for Latvian language

Askars Salimbajevs; Jevgenijs Strigins

Error analysis and improving speech recognition for Latvian language

Askars Salimbajevs
, Jevgenijs Strigins

Tilde Company

Zinātniskās darbības rezultāts: Devums žurnālam › Konferences zinātniskais raksts › koleģiāli recenzēts

5 Atsauces (Scopus)

Kopsavilkums

Developing a large vocabulary automaticspeech recognition system is a very difficult task, due to the high variations in domain andacoustic variability. This task is even more difficult for the Latvian language, which is veryrich morphologically and in which one word can have dozens of surface forms. Althoughthere is some research on speech recognition for Latvian, Latvian ASR remains behind "big" languages such as English, German etc. In order to improve the performance of LatvianASR, it is important to understand what errors does it make and why. In this paper, the authorsanalyze the most common errors of Latvian ASR. Based on this, baseline system WER isimproved from 30.94% to 28.43%.

Oriģinālvaloda	Angļu
Lapas (no-līdz)	563-569
Lapu skaits	7
Žurnāls	International Conference Recent Advances in Natural Language Processing, RANLP
Sējums	2015-January
Publikācijas statuss	Publicēts - 2015
Ārēji publicēts	Jā
Pasākums	10th International Conference on Recent Advances in Natural Language Processing, RANLP 2015 - Hissar, Bulgārija Ilgums: 7 sept. 2015 → 9 sept. 2015

OECD Zinātnes nozare

1.2 Datorzinātne un informātika

Citi faili un saites

Citēt šo

@article{213c01de813a413f92a62cc8310f641b,

title = "Error analysis and improving speech recognition for Latvian language",

abstract = "Developing a large vocabulary automaticspeech recognition system is a very difficult task, due to the high variations in domain andacoustic variability. This task is even more difficult for the Latvian language, which is veryrich morphologically and in which one word can have dozens of surface forms. Althoughthere is some research on speech recognition for Latvian, Latvian ASR remains behind {"}big{"} languages such as English, German etc. In order to improve the performance of LatvianASR, it is important to understand what errors does it make and why. In this paper, the authorsanalyze the most common errors of Latvian ASR. Based on this, baseline system WER isimproved from 30.94\% to 28.43\%.",

author = "Askars Salimbajevs and Jevgenijs Strigins",

year = "2015",

language = "English",

volume = "2015-January",

pages = "563--569",

journal = "International Conference Recent Advances in Natural Language Processing, RANLP",

issn = "1313-8502",

publisher = "Incoma Ltd",

note = "10th International Conference on Recent Advances in Natural Language Processing, RANLP 2015 ; Conference date: 07-09-2015 Through 09-09-2015",

}

TY - JOUR

T1 - Error analysis and improving speech recognition for Latvian language

AU - Salimbajevs, Askars

AU - Strigins, Jevgenijs

PY - 2015

Y1 - 2015

N2 - Developing a large vocabulary automaticspeech recognition system is a very difficult task, due to the high variations in domain andacoustic variability. This task is even more difficult for the Latvian language, which is veryrich morphologically and in which one word can have dozens of surface forms. Althoughthere is some research on speech recognition for Latvian, Latvian ASR remains behind "big" languages such as English, German etc. In order to improve the performance of LatvianASR, it is important to understand what errors does it make and why. In this paper, the authorsanalyze the most common errors of Latvian ASR. Based on this, baseline system WER isimproved from 30.94% to 28.43%.

AB - Developing a large vocabulary automaticspeech recognition system is a very difficult task, due to the high variations in domain andacoustic variability. This task is even more difficult for the Latvian language, which is veryrich morphologically and in which one word can have dozens of surface forms. Althoughthere is some research on speech recognition for Latvian, Latvian ASR remains behind "big" languages such as English, German etc. In order to improve the performance of LatvianASR, it is important to understand what errors does it make and why. In this paper, the authorsanalyze the most common errors of Latvian ASR. Based on this, baseline system WER isimproved from 30.94% to 28.43%.

UR - https://www.scopus.com/pages/publications/84949795828

UR - http://lml.bas.bg/ranlp2015/docs/RANLP_main.pdf

M3 - Conference article

AN - SCOPUS:84949795828

SN - 1313-8502

VL - 2015-January

SP - 563

EP - 569

JO - International Conference Recent Advances in Natural Language Processing, RANLP

JF - International Conference Recent Advances in Natural Language Processing, RANLP

T2 - 10th International Conference on Recent Advances in Natural Language Processing, RANLP 2015

Y2 - 7 September 2015 through 9 September 2015

ER -

Error analysis and improving speech recognition for Latvian language

Kopsavilkums

OECD Zinātnes nozare

Citi faili un saites

Nospiedums

Citēt šo