“(Interspeech 2021 Special Session)”版本间的差异
（以“==Interspeech 2021 OLR Special Session== Oriental languages are rich and complex. In general, these languages can be grouped into several language families, and ea...”为内容创建页面） |
|||
| 第8行： | 第8行： | ||
OLR 2020 sets up three tasks as follows: | OLR 2020 sets up three tasks as follows: | ||
| − | + | *Task 1: cross-channel recognition: A close-set identification task, which means the language of each utterance is among the known traditional 6 target languages, but utterances were recorded with different channels. | |
| − | + | *Task 2: dialect identification: An open-set identification task, in which three nontarget languages are added to the test set with the three target dialects. | |
| − | + | *Task 3: noisy recognition: A close-set identification in noisy conditions. | |
The top performance of the participants is very promising. They proposed novel and interesting techniques from various aspects, including feature extraction, embedding architecture, back-end modeling, normalization techniques. Some of them seem highly promising. | The top performance of the participants is very promising. They proposed novel and interesting techniques from various aspects, including feature extraction, embedding architecture, back-end modeling, normalization techniques. Some of them seem highly promising. | ||
2021年1月18日 (一) 07:52的版本
Interspeech 2021 OLR Special Session
Oriental languages are rich and complex. In general, these languages can be grouped into several language families, and each family involves many diverse languages. With such diversity, oriental language is a treasure for multilingual research.
Language recognition (LR) is one of the key tasks for multilingual speech processing. In 2016, Tsinghua University and Speech Ocean initialized the Oriental Language Recognition (OLR) challenge, with aim to promote the LR research in particular for oriental languages that are limited in source and are difficult to be distinguished. Since then, OLR challenge has been conducted for 5 years, with more and more impact. In 2020, we had 90 participants from all over the world.
OLR 2020 sets up three tasks as follows:
- Task 1: cross-channel recognition: A close-set identification task, which means the language of each utterance is among the known traditional 6 target languages, but utterances were recorded with different channels.
- Task 2: dialect identification: An open-set identification task, in which three nontarget languages are added to the test set with the three target dialects.
- Task 3: noisy recognition: A close-set identification in noisy conditions.
The top performance of the participants is very promising. They proposed novel and interesting techniques from various aspects, including feature extraction, embedding architecture, back-end modeling, normalization techniques. Some of them seem highly promising.
In order to summarize the technical advance, we propose a special session on Interspeech 2021. The special session will be a forum to exchange ideas regarding language recognition under very challenging conditions, e.g., cross channel, low resource and complex acoustic ambient. All the topics are based on the OLR20 protocol, so a special session will be the best.
The proposed session will be in the form of oral presentation. If possible, we also hope to have some time for panel discussion, mainly for the next OLR challenge.
