Datasets:

MushanW
/

GLOBE_V3

language:
  - en
license: cc0-1.0
source_datasets:
  - mozilla-foundation/common_voice_14_0
task_categories:
  - text-to-audio
  - automatic-speech-recognition
  - audio-to-audio
  - audio-classification
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
      - split: dev
        path: data/dev-*
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: snr
      dtype: float32
    - name: whisper_transcription_large_v3
      dtype: string
    - name: utmos
      dtype: float32
    - name: wer
      dtype: float32
    - name: cer
      dtype: float32
    - name: predicted_gender
      dtype: string
    - name: predicted_accent
      dtype: string
    - name: predicted_age
      dtype: string
    - name: common_voice_path
      dtype: string
    - name: common_voice_sentence_id
      dtype: string
    - name: common_voice_sentence
      dtype: string
    - name: common_voice_age
      dtype: string
    - name: common_voice_gender
      dtype: string
    - name: common_voice_accents
      dtype: string
    - name: audio
      dtype:
        audio:
          sampling_rate: 24000
  splits:
    - name: test
      num_bytes: 2016767055.913
      num_examples: 9179
    - name: dev
      num_bytes: 2212771887.685
      num_examples: 9603
    - name: train
      num_bytes: 180237198132.75
      num_examples: 704750
  download_size: 165538220540
  dataset_size: 184466737076.348

Important notice

Differences between V3 version and two previous versions (V1|V2):

This version is built base on Common Voice 21.0 English Subset.
This version only includes utterance that are an exact match with the transcription from Whisper V3 LARGE (CER == 0).
This version includes the original Common Voice metadata (age, gender, accent, and ID).
All audio files in this version are at 24kHz sampling rate.
All audio files in this version are unenhanced. (We’d greatly appreciate it if anyone is willing to provide an API for speech enhancement.)

Globe

The full paper can be accessed here: arXiv

An online demo can be accessed here: Github

Abstract

This paper introduces GLOBE, a high-quality English corpus with worldwide accents, specifically designed to address the limitations of current zero-shot speaker adaptive Text-to-Speech (TTS) systems that exhibit poor generalizability in adapting to speakers with accents. Compared to commonly used English corpora, such as LibriTTS and VCTK, GLOBE is unique in its inclusion of utterances from 23,519 speakers and covers 164 accents worldwide, along with detailed metadata for these speakers. Compared to its original corpus, i.e., Common Voice, GLOBE significantly improves the quality of the speech data through rigorous filtering and enhancement processes, while also populating all missing speaker metadata. The final curated GLOBE corpus includes 535 hours of speech data at a 24 kHz sampling rate. Our benchmark results indicate that the speaker adaptive TTS model trained on the GLOBE corpus can synthesize speech with better speaker similarity and comparable naturalness than that trained on other popular corpora. We will release GLOBE publicly after acceptance.

Citation

@misc{wang2024globe,
      title={GLOBE: A High-quality English Corpus with Global Accents for Zero-shot Speaker Adaptive Text-to-Speech}, 
      author={Wenbin Wang and Yang Song and Sanjay Jha},
      year={2024},
      eprint={2406.14875},
      archivePrefix={arXiv},
}