BEIJING, July 4, 2022 /PRNewswire/ -- Magic Data's paper 'Open Source MagicData-RAMC: A Rich Annotated Mandarin Conversational(RAMC) Speech Dataset' is accepted by INTERSPEECH 2022, the world's largest and most comprehensive conference on the science and technology of spoken language processing. Themed "Human and Humanizing Speech Technology", INTERSPEECH 2022 will take place from September 18-22 virtually and in Incheon Korea.
In order to further enrich the open source speech corpus and promote the development of speech language processing technology and conversational AI, this year in April, Magic Data officially released a 180-hour Chinese conversational speech dataset — MagicData-RAMC on MagicHub, an open source community, developed by Magic Data.
The Paper introduces MagicData-RAMC dataset and research based on the dataset conducted by Magic Data, together with the Institute of Acoustics, Chinese Academy of Sciences, Shanghai Jiao Tong University and Northwestern Polytechnic University.
As a collection of high quality and richly annotated training data, MagicData-RAMC includes 351 sets of multi-turn Mandarin conversations recorded in indoor environment by smart phone with a total duration of 180 hours. The data set can support developers in completing research on speech recognition, speaker diarization, and keyword search.
For the Automatic Speech Recognition task, the research team use a Conformer-based end-to-end (E2E) model implemented by ESPnet2 toolkit. The experimental result is shown in terms of character error rate (CER) as 19.1.
For speaker diarization task, the baseline system consists of three components: speaker activity detection (SAD), speaker embedding extractor and clustering. Experimental result shows that JER is 47.49 on test set.
The keyword search task follows the DTA Att-E2E-KWS approach proposed in relying on attention-based E2E ASR framework and frame-synchronous phoneme alignments. The result shows that the system got 0.8587 on precision rate.
Paper available on arxiv
About Magic Data
Magic Data is a global AI data solutions provider headquartered in Beijing, providing professional data solutions to enterprises and academic institutions engaged in artificial intelligence R&D and application research to voice recognition (ASR), speech synthesis (TTS), natural language processing (NLP), and computer vision (CV).
MagicHub community is an open-source data platform developed by Magic Data dedicated to assisting AI developers in model training and to promoting the development of an open-source ecosystem.