Implementation of a Radiology Speech Recognition System for Estonian Using Open Source Software

Abstract

Speech recognition has become increasingly popular in radiology reporting in the last decade. However, developing a speech recognition system for a new language in a highly specific domain requires a lot of resources, expert knowledge and skills. Therefore, commercial vendors do not offer ready-made radiology speech recognition systems for less-resourced languages.This paper describes the implementation of a radiology speech recognition system for Estonian, a language with less than one million native speakers. The system was developed in partnership with a hospital that provided a corpus of written reports for language modeling purposes. Rewrite rules for pre-processing training texts and postprocessing recognition results were created manually based on a small parallel corpus created by the hospital's radiologists, using the Thrax toolkit. Deep neural network based acoustic models were trained based on 216 hours of out-of-domain data and adapted on 14 hours of spoken radiology data, using the Kaldi toolkit. The current word error rate of the system is 5.4%. The system is in active use in real clinical environment.

Preschool and primary teachers’ evaluations of their preparation, competence and experiences to teach newly arrived migrant students

“…Intervjuusid analüüsiti induktiivse sisuanalüüsi meetodil, mis võimaldab saada ülevaate tekstist kui tervikust, kirjeldada süstemaatiliselt kvalitatiivsete andmete tähendust ning analüüsida ridade vahele peidetut (Schreier, 2012). Transkri beerimisel kasutati automaatse kõnetuvastuse veebipõhist tarkvara (Alumäe et al, 2018), seejärel korrigeeris esimene autor tekste kuulmise järgi. Trans kribeeritud teksti ühe intervjuu kohta oli keskmiselt 20 lehekülge (kokku 275 lehekülge).…”

Võõrkeele omandamise kogemused eestlaste mälestustes ja arvamustes

“…Intervjuude transkribeerimisel on tarvitatud eesti keele suulise kõne üleskirjutamisel kasutatavaid printsiipe (vt nt Hennoste et al 2009 ja tingmärkide seletust artikli lõpus) ja veebipõhise kõnetuvas tuse abi (Alumäe et al 2018). Litereeringud kuulas üle ja täpsustas magistrant Saara Liis Jõerand.…”

Towards Accurate Text Verbalization for ASR Based on Audio Alignment

“…On the other hand, context-independent rules select a single verbalization variant for each non-lexical unit which is usually very inaccurate because of the aforementioned ambiguities of written language. Alumäe et al (2017) show how a small amount of verbalized text that serves as ground truth can be used to mitigate the lack of verbalization variability when using context-independent rules. Their method chooses a verbalization for each sentence by sampling from all of its possible verbalization variants with probability that is proportional to the probability of each individual variant.…”