Dataocean AI Has Participated in Creating the Open-Source Dataset GigaSpeech 2: A Large-Scale and Multi-Domain ASR Corpus for Low-Resource Languages

Dataocean AI has collaborated with Shanghai Jiao Tong University, The Chinese University of Hong Kong, Tsinghua University, Pengcheng Lab, AISpeech, Birch AI, and Seasalt AI to successfully develop GigaSpeech 2. The development and test sets of GigaSpeech 2 are labeled by a professional team from Dataocean AI.

This press release features multimedia. View the full release here: https://www.businesswire.com/news/home/20240924609911/en/

GigaSpeech 2 Overview

GigaSpeech 2 is an ever-expanding, large-scale, multi-domain, and multilingual speech recognition corpus designed to promote research and development in low-resource language speech recognition. GigaSpeech 2 raw contains 30,000 hours of automatically transcribed audio, covering Thai, Indonesian, and Vietnamese. After multiple rounds of refinement and iteration, GigaSpeech 2 refined offers 10,000 hours of Thai, 6,000 hours of Indonesian, and 6,000 hours of Vietnamese. The test sets labeled by Dataocean AI for Thai and Indonesian, each consist of 10 hours, while the development sets are 10 hours for Thai and Indonesian. The team have also open-sourced multilingual speech recognition models trained on the GigaSpeech 2 data, achieving performance comparable to commercial speech recognition services.

Dataset Construction

The construction process of GigaSpeech 2 has also been open-sourced. This is an automated process for building large-scale speech recognition datasets from vast amounts of unlabeled audio available on the internet. The automated process involves data crawling, transcription, alignment, and refinement. Initially, Whisper is used for preliminary transcription, followed by forced alignment with TorchAudio to produce GigaSpeech 2 raw through multi-dimensional filtering. The dataset is then refined iteratively using an improved Noisy Student Training (NST) method, enhancing the quality of pseudo-labels through repeated iterations, ultimately resulting in GigaSpeech 2 refined.

GigaSpeech 2 encompasses a wide range of thematic domains, including agriculture, art, business, climate, culture, economics, education, entertainment, health, history, literature, music, politics, relationships, shopping, society, sports, technology, and travel. Additionally, it covers various content formats such as audiobooks, documentaries, lectures, monologues, movies and TV shows, news, interviews, and video blogs.

Training Set Details

GigaSpeech 2 offers a comprehensive and diverse training set, which is meticulously designed to support the development of robust and high-performing speech recognition models. The training set details are as follows:

- Thai: The raw version consists of 12,901.8 hours of speech data, while the refined version encompasses 10,262.0 hours.
- Indonesian: The raw data amounts to 8,112.9 hours, and the refined data comprises 5,714.0 hours.
- Vietnamese: The raw dataset includes 7,324.0 hours of speech recordings, with the refined dataset totaling 6,039.0 hours.

Development and Test Set Details

Dataocean AI’s COO - Ke Li, who is also one of the paper's authors, has led GigaSpeech 2 test sets project. With nearly 20 years of project experience, the team has contributed in Thai and Indonesian with word accuracy of over 97%. Besides those two East Asian languages, Dataocean AI’s team can also cover over 200 languages and dialects around the world. The company offer 1600+ high-quality off-the-shelf datasets are applicable for multiple scenarios such as Generative AI, Autonomous driving, Smart home, Customer services and etc., fulfilling the evolving needs of the AI industry.

Experimental Results

We conducted a comparative evaluation of speech recognition models trained on the GigaSpeech 2 dataset against industry-leading models, including OpenAI Whisper (large-v3, large-v2, base), Meta MMS L1107, Azure Speech CLI 1.37.0, and Google USM Chirp v2. The comparison was carried out in Thai, Indonesian, and Vietnamese languages. Performance evaluation was based on three test sets: GigaSpeech 2, Common Voice 17.0, and FLEURS, using Character Error Rate (CER) or Word Error Rate (WER) as metrics. The results indicate:

Thai: Our model demonstrated exceptional performance, surpassing all competitors, including commercial interfaces from Microsoft and Google. Notably, our model achieved this significant result while having only one-tenth the number of parameters compared to Whisper large-v3.

Indonesian and Vietnamese: Our system exhibited competitive performance compared to existing baseline models in both Indonesian and Vietnamese languages.

Resource Links

The GigaSpeech 2 dataset is now available for download:
https://huggingface.co/datasets/speechcolab/gigaspeech2

The automated process for constructing large-scale speech recognition datasets is available at:
https://github.com/SpeechColab/GigaSpeech2

The preprint paper is available at:
https://arxiv.org/pdf/2406.11546

Dataocean AI website:
https://www.dataoceanai.com

View source version on businesswire.com: https://www.businesswire.com/news/home/20240924609911/en/