Article

Open-Source End-To-End Learning for Privacy-Preserving German ASR (online)

* Presenting author
Day / Time: 23.03.2022, 15:00-15:20
Room: 57-02
Typ: Regulärer Vortrag
Online-access: Bitte loggen Sie sich ein, damit weitere Inhalte sichtbar werden (bspw. der Zugang zur Onlinesitzung).
Abstract:

As part of HUMAINE (HUMan centered AI NEtwork), we are developing a privacy-preserving automatic speech recognition (ASR) system for the use case of automated, structured documentation of nurse-patient interviews.This paper firstly describes the training and evaluation for our German-language end-to-end model. It is based on a convolutional recurrent DNN model and optimized using sequence-to-sequence learning.Utilizing SpeechBrain in training, we achieve a word error rate below 9% for an open-vocabulary task. The small memory footprint of our model allows for the recognizer to be executed on a desktop computer, enabling privacy-preserving, purely local speech transcription.Secondly, we describe our initial natural language processing (NLP) experiments for transforming text to structured information, based on an own labeled dialogue dataset, covering five thematic question-answer pairs of every-day dialogues. We evaluated different machine learning algorithms for grouping dialogue snippets to topics. Overall, a random forest model yielded the best classification accuracy of about 90%. Next, supported by health experts, we will create a database of realistic nurse-patient dialogues, containing information to fill in a specific professional patient admission form, to evaluate our NLP model.Within this contribution, we detail the concept and performance of our ASR and NLP system.

Error

Error : E_DEPRECATED

strpos(): Passing null to parameter #1 ($haystack) of type string is deprecated

1
/kunden/431424_13355/webseiten/subdomains/app2022.daga-tagung.de/memyc/resources/OutputCompression.php, line 35
32.              */
33.             if (headers_sent()) {
34.                 $encoding = false;
35.             } elseif (strpos($HTTP_ACCEPT_ENCODING, 'x-gzip') !== false) {
36.                 $encoding = 'x-gzip';
37.             } elseif (strpos($HTTP_ACCEPT_ENCODING, 'gzip') !== false) {
38.                 $encoding = 'gzip';