Online Punctuation Restoration using ELECTRA Model for streaming ASR Systems

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F46747885%3A24220%2F23%3A00011545" target="_blank" >RIV/46747885:24220/23:00011545 - isvavai.cz</a>

Result on the web

<a href="https://www.isca-archive.org/interspeech_2023/polacek23_interspeech.html" target="_blank" >https://www.isca-archive.org/interspeech_2023/polacek23_interspeech.html</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.21437/Interspeech.2023-664" target="_blank" >10.21437/Interspeech.2023-664</a>

Result language

angličtina

Original language name

Online Punctuation Restoration using ELECTRA Model for streaming ASR Systems

Original language description

In this work, we propose a lightweight online approach to automatic punctuation restoration (APR), which can be utilized in speech transcription systems for, e.g., live captioning TV or radio streams. It uses only text input without prosodic features and a fine-tuned ELECTRA-Small model with a two-layer classification head. It allows for restoring question marks, commas, and periods with a very short inference time and a low latency of just three words. Our APR scheme is first tuned and compared to other architectures on a set of manual TV news transcripts. The resulting system is then compared to another real-time APR module utilizing a recurrent network and a combination of text and acoustic features. The test data we use contains automatic transcripts of radio talks and TV debates; we are also publishing this data. The results show that our APR module performs better than the above-mentioned system and yields on the two test sets an average F1 of 71.2% and 69.4%, respectively.

Czech name

—

Czech description

—

Type

D - Article in proceedings

CEP classification

—

OECD FORD branch

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Project

<a href="/en/project/TO01000027" target="_blank" >TO01000027: NORDTRANS - Technology for automatic speech transcription in selected Nordic languages</a><br>

Continuities

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Publication year

2023

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Article name in the collection

Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH

ISBN

—

ISSN

2308-457X

e-ISSN

—

Number of pages

5

Pages from-to

446-450

Publisher name

ISCA

Place of publication

Irsko

Event location

Dublin, Irsko

Event date

Jan 1, 2023

Type of event by nationality

WRD - Celosvětová akce

UT code for WoS article

—