GigaSpeech

A Large, modern and evolving dataset for automatic speech recognition.

Audio Source	Hours	acoustic	domain
Podcast	3,498	near-field, indoor, spontaneous, background music, genders, ages, accents	daily topics
Youtube	3,845	near & far field, indoor & outdoor, reading & spontaneous, clean & loud background music/noises, genders, ages, accents	vast topics/domains coverage
Audiobook	2,655	slow, clear, readings	books, stories
total	10,000

Training Set

We organize the entire dataset via 5 subsets, targeting on different users.

Subset	Notation	Size(Hours)	Target Usage
eXtra Small	XS	10	pipeline/recipe coding & debugging
Small	S	250	quick research experiment for new ideas
Medium	M	1000	serious research experiment / quick industrial experiment
Large	L	2500	serious industrial experiment
eXtra Large	XL	10000	industrial system building

{XL includes {L includes {M includes {S includes {XS}}}}}

Dev & Test Sets

Dev set

• source:
  1. random drawn audios from crawled podcast & youtube data

Test set

• source:
  1. random drawn audios from crawled podcast & youtube data
  2. independent to crawling process, GigaSpeech authors manually collected podcasts & videos from internet, to cover wider scenarios & domains.

(Dev + Test) sets contain 50~60 hours raw data, labeled by payed professional human annotators.

Dataset download

To download the dataset, do the following steps:

1. Put aliyun_ossutil.cfg in SAFEBOX folder

2. Run the following steps for downloading the dataset only

   utils/gigaspeech_download.sh /download/destination/dir/for/GigaSpeechDataset

   Then the entire dataset will be downloaded to your local dir.

   If your network is interrupted or broken during downloading, you can just rerun above command, it will continue with previous downloading.

   You can also use above command to update your local GigaSpeech copy with newest GigaSpeech release.

Toolkit Support

We maintain data preparation scripts for different speech recognition toolkits in this repository so that when we update the dataset (note, this is an evolving dataset), we don't have to update the scripts in the downstream toolkits. Data preparation scripts for different speech recognition toolkits are maintained in the toolkits/ folder, e.g., toolkits/kaldi for the Kaldi speech recognition toolkit.

Data Preparation for Toolkits

To use the data preparation scripts, do the following in your toolkit (here we use Kaldi as an example)

git clone https://github.com/SpeechColab/GigaSpeech.git

cd GigaSpeech
utils/gigaspeech_download.sh /disk1/audio_data/gigaspeech
toolkits/kaldi/gigaspeech_data_prep.sh --train-subset XL /disk1/audio_data/gigaspeech ../data
cd ..

Notes on Text Processing

1. By design we have punctuations in labels. To be specific, 4 punctuations may appear in utterance's text_tn section, they are:

   <COMMA>
   <PERIOD>
   <QUESTIONMARK>
   <EXCLAMATIONPOINT>
   

2. Grabage utterance tags in DEV/TEST sets: our DEV/TEST sets are labelled by human annotators, they are instructed to label every single piece of the entire audio. So if part of audio is not human speech, they label it with a set of garbage utterance tags. A complete table of garbage meta tags are listed below:

   <SIL>
   <MUSIC>
   <NOISE>
   <OTHER>
   

   utterances with these garbage tags are not considered to be valid speech. So our recommendation is to discard these utterances in downstream training/testing. The reason why we keep these tags is to keep the integrity of human labels, so there is no "gap" inside DEV/TEST labels.

Add Support for a New Toolkit

To add data preparation support for a new toolkit, please follow toolkits/kaldi/gigaspeech_data_prep.shand add similar scripts for your own toolkit. For example, for ESPnet2, you would add toolkits/espnet2/gigaspeech_data_prep.sh to prepare the dataset, and all other related scripts should be maintained under toolkits/espnet2.