Tạo phụ đề video dựa trên kỹ thuật nhận dạng giọng nói: Thử nghiệm cho một số chương trình tại VTV

0 Người đánh giá. Xếp hạng trung bình 0

Tác giả: Hữu Phong Nguyễn, Minh Trung Trần, Nguyễn Quốc Bảo Võ

Ngôn ngữ: vie

Ký hiệu phân loại: 006.45 *Acoustical pattern recognition

Thông tin xuất bản: Tạp chí Khoa học Giáo dục Kỹ thuật - Trường Đại học Sư phạm Kỹ thuật TP. Hồ Chí Minh, 2022

Mô tả vật lý: 38-48

Bộ sưu tập: Metadata

ID: 499494

Thêm vào giỏ

Bài báo này trình bày kết quả thử nghiệm công cụ nhận dạng giọng nói Speech-To-Text (STT) cho các nội dung VOD (Video On Demand) trên hệ thống VTVgo của Đài THVN. Để đánh giá độ chính xác của công cụ STT, tỷ lệ lỗi từ (WER: Word Error Rate) được sử dụng để đo hiệu suất của hệ thống nhận dạng giọng nói tự động, dịch máy. Kết quả thử nghiệm thực hiện 10 thể loại chương trình truyền hình khác nhau với 1065 giờ video. Tỉ lệ WER thấp nhất là 2.8% đến 4.3% đạt được với một số thể loại chương trình thời sự và tin tức, dự báo thời tiết, ở đó phần lớn người nói, người dẫn chương trình (MC) đọc giọng chuẩn trong Studio và lời thoại từ một người nói, ít bị nhiễu bởi tạp âm bên ngoài. Bên cạnh đó, để minh họa ứng dụng phụ đề video, chúng tôi tiến hành thử nghiệm trên hệ thống VTVgo, tích hợp công cụ hiển thị phụ đề tùy chọn vào ứng dụng VTVgo app. Nền tảng thử nghiệm là SmartTV và SmartPhone Android, nhằm minh họa khả năng ứng dụng phụ đề video trên nền tảng phân phối nội dung số OTT (Over The Top).This paper presents the trial results of Speech-To-Text (STT) recognition tool for VOD (Video On Demand) contents of the VTVgo system at Vietnam Television. In order to evaluate the accuracy of the STT tool, the word error rate (WER) was used to measuring the performance of the automatic speech recognition, the machine translation system. Test results of 10 different types of TV show with 1065 video hours were analyzed. The WER had achieved low level from 2.8% to 4.3% with some genres of news, 19h, weather forecasts, where the majority of speakers, presenters (MC) read standard voices in the Studio. The dialogue from a speaker, less interference from outside noise. Besides, to illustrating the video subtitle application, we had conducted the test on the VTVgo system, integrated the optional subtitle display tool into the VTVgo app. The test Android platform was Smart TV and SmartPhone, to demonstrating the ability to apply video subtitles on the OTT (Over The Top) - the digital content distribution platform.

1. Đọc trực tuyến

1. Đọc trực tuyến

Tạo bộ sưu tập với mã QR