GigaSpeech

This is the official repository of the GigaSpeech dataset. For details of how we created the dataset, please refer to our Interspeech paper: "GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio". Preprint available on arxiv.

GigaSpeech version: 1.0.0 (07/05/2021)

Download

1. Step 1: Please fill out the Google Form here
2. Step 2:
   • Option A: Follow the instructions in replied email from SpeechColab to get the raw release of GigaSpeech
   • Option B: Refer to GigaSpeech On HuggingFace to get a pre-processed version of GigaSpeech via HuggingFace.

Leaderboard

Contributor	Toolkit	Train Recipe	Train Data	Inference	Dev/Test WER
Baseline	Athena	Transformer-AED + RNNLM	GigaSpeech v1.0.0 XL	model example	13.60 / 12.70
Baseline	Espnet	Conformer/Transformer-AED	GigaSpeech v1.0.0 XL	model example	10.90 / 10.80
Baseline	Kaldi	Chain + RNNLM	GigaSpeech v1.0.0 XL	model example	14.78 / 14.84
Baseline	Pika	RNN-T	GigaSpeech v1.0.0 XL	model example	12.30 / 12.30
Johns Hopkins University	Icefall	Transducer: Zipformer encoder + Embedding decoder	GigaSpeech v1.0.0 XL	model example	10.25 / 10.38
Johns Hopkins University	Icefall	Pruned Stateless RNN-T	GigaSpeech v1.0.0 XL	model example	10.40 / 10.51
Johns Hopkins University	Icefall	Conformer CTC + ngram & attention rescoring	GigaSpeech v1.0.0 XL	model example	10.47 / 10.58
Mobvoi	Wenet	Joint CTC/AED(U2++)	GigaSpeech v1.0.0 XL	model example	10.70 / 10.60
ByteDance AI Lab	NeurST	Transformer-AED	GigaSpeech v1.0.0 XL	model example	11.89 / 11.60

Dataset

Audio Source

• Language: English
• 33,000+ hours for unsupervised/semi-supervised learning
• 10,000 hours with high-quality human transcriptions for supervised learning

Audio Source	Transcribed Hours	Total Hours	Acoustic Condition
Audiobook	2,655	11,982	Reading Various ages and accents
Podcast	3,498	9,254	Clean or background music Indoor Near-field Spontaneous Various ages and accents
YouTube	3,845	11,768	Clean and noisy Indoor and outdoor Near- and far-field Reading and spontaneous Various ages and accents
total	10,000	33,005

Transcribed Training Subsets

Subset	Hours	Remarks
XS	10	System building and debugging
S	250	Quick research experiments
M	1,000	Large-scale research experiments
L	2,500	Medium-scale industrial experiments
XL	10,000	Large-scale industrial experiments

Larger subsets are supersets of smaller subsets, e.g., subset L contains all the data from subset M.

Transcribed Evaluation Subsets

Subset	Hours	Remarks
Dev	12	Randomly selected from the crawled Podcast and YouTube Data
Test	40	Part of the subset was randomly selected from the crawled Podcast and YouTube data; part of it was manually collected through other channels to have better coverage.

Evaluation subsets are annotated by professional human annotators

Data Preparation Guidelines

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.

Preparation Scripts

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/download_gigaspeech.sh /disk1/audio_data/gigaspeech
toolkits/kaldi/gigaspeech_data_prep.sh --train-subset XL /disk1/audio_data/gigaspeech ../data
cd ..

Metadata walkthrough

We save all the metadata information to a single JSON file named GigaSpeech.json. Below is a snip of this file:

{
  "dataset": "GigaSpeech",
  "language": "EN",
  "version": "v1.0.0",
  ... ...
  "audios": [
    {
      "title": "The Architect of Hollywood",
      "url": "https://99percentinvisible.org/episode/the-architect-of-hollywood/download",
      "path": "audio/podcast/P0001/POD0000000025.opus",
      ... ...
      "segments": [
        {
          "sid": "POD0000000025_S0000103",
          "speaker": "N/A",
          "begin_time": 780.31,
          "end_time": 783.13,
          "text_tn": "FOUR O'CLOCK TOMORROW AFTERNOON <COMMA> SAID WILLIAMS <PERIOD>",
          "subsets": [
            "{XL}",
            "{L}"
          ]
        },
        ... ...
      ],
      ... ...
    },
    ... ...
  ]
}

To use the corpus, users are expected to extract the relevant information from GigaSpeech.json. For example, for the speech recognition task, one should first follow the "audios" entry, and work out a list of audio files. One can then follow the "url" entry to download the original audio file, or "path" if preprocessed audio files have been downloaded to the disk. After that, for each audio file, one can follow the "segments" entry, and work out the trainable audio segments, as well as their corresponding transcripts. Of course, we also have various supplementary entries, such as "subsets", "md5", which will also be helpful for your task.

The metadata file GigaSpeech.json is version controlled, and is supposed to get updated over the time. In future releases, we plan to add speaker information to the metadata file, so that it will be suitable for speaker identification/verification tasks. We also plan to add more data from different sources to increase the diversity.

We also provide some convenient command-line tools based on jq, e.g., utils/ls_audio.sh, utils/show_segment_info.sh, utils/ls_md5.sh.

Audio Processing

• Resampling: GigaSpeech audio files are resampled at 16 kHz sampling rate, and are compressed with the Opus format. The Opus compression, however, does not depend on the input sample rate; it uses the bandwidth instead. Timestamps are measured in 48 kHz units even if the full bandwidth is not used. Likewise, the output sample rate may be freely chosen. For example, audio can be input at 16 kHz yet be set to encode only narrowband audio. For this reason, we recommend our users to explicitly resample the decoded audio to 16 kHz sampling rate before training & testing. For opus-to-wav conversion, refer to our exampler tool utils/opus_to_wav.py

Text Pre-Processing

• Punctuations: We keep 4 punctuations in the normalized text (see the text_tn entry in GigaSpeech.json)

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

  This allows researchers to explore directions such as end-to-end endpointing and punctuation restoration. If you don't need these, you can remove them for your own training.

• Garbage Utterance Tags: The Dev/Test evaluation sets are annotated by human annotators. They are instructed to label the entire audio file without "gaps". So for non-speech segments, garbage utterance tags are used instead. We recommend our users to discard these utterances in your training. A complete list of these tags are:

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

Text Post-Processing (before scoring)

• Conversational Fillers: Spontaneous/Conversational speech contains conversational fillers such as:

  'UH', 'UHH', 'UM', 'EH', 'MM', 'HM', 'AH', 'HUH', 'HA', 'ER'
  

  We recommend our users to remove these fillers from both hypothese and reference text before WER scoring, so that we will have apple-to-apple performance comparisons across different toolkits. See discussion on post-processing here. We also provide a scoring tool utils/gigaspeech_scoring.py and this tool is used by all the toolkits reported in above leaderboard section.

Add Support for a New Toolkit

To add data preparation support for a new toolkit, please follow toolkits/kaldi/gigaspeech_data_prep.sh and 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.

Collaboration

We are a group of volunteers trying to make speech technologies easier to use. We welcome any kind of contributions. Currently we are exploring the following directions. If you are interested in one of the directions, and you think you will be able to help, please contact [email protected].

• Inference architecture for different pre-trained models
• Adding diverse audio source
• Benchmarking speech algorithms/services
• Building and releasing pre-trained models
• Supporting more languages
• Supporting more tasks through GigaSpeech.json (e.g., speaker ID)
• Making new datasets with permissive licenses

Institutional Contributors

Institution	Contribution
IEIT, Tsinghua University	Computing power; Data host; Researchers
Magic Data	Data host mirror
speechocean	Data host mirror; Evaluation data annotation
Xiaomi Corporation	Computing power; Researchers

Citation

Please cite our paper if you find this work useful:

@inproceedings{GigaSpeech2021,
  title={GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio},
  booktitle={Proc. Interspeech 2021},
  year=2021,
  author={Guoguo Chen, Shuzhou Chai, Guanbo Wang, Jiayu Du, Wei-Qiang Zhang, Chao Weng, Dan Su, Daniel Povey, Jan Trmal, Junbo Zhang, Mingjie Jin, Sanjeev Khudanpur, Shinji Watanabe, Shuaijiang Zhao, Wei Zou, Xiangang Li, Xuchen Yao, Yongqing Wang, Yujun Wang, Zhao You, Zhiyong Yan}
}

Contact

If you have any concerns, please contact [email protected].

Metadata Changelog

• 07/23/2021 v1.0.0: We found a bug in the metadata and fixed that. We made an exception and kept the version number the same because this correct version was used in the original experiments in the paper.
• 07/05/2021 v1.0.0: Initial release.