Hugging Face's logo

Datasets:
SpeechAntiSpoofingBenchmarks
/
HABLA

Tasks:
Audio Classification
Modalities:
Audio
Text
Formats:
parquet
Languages:
Spanish
Size:
10K - 100K
Tags:
anti-spoofing
audio-deepfake-detection
speech
benchmark
arena-ready
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
19
HABLA / README.md
korallll's picture
korallll
Add maintainer contact: email + Telegram channel
7e18777 verified
preview code
|
Raw
History Blame Contribute Delete
3.92 kB
metadata 
license: cc-by-4.0
language:
  - es
pretty_name: HABLA
task_categories:
  - audio-classification
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet
tags:
  - anti-spoofing
  - audio-deepfake-detection
  - speech
  - benchmark
  - arena-ready
arxiv:
  - 10.21437/Interspeech.2023-2272

HABLA

Benchmark-ready packaging of HABLA — HABLA: A Dataset of Latin American Spanish Accents for Voice Anti-spoofing (Tamayo Flórez, Manrique & Pereira Nunes, Interspeech 2023). The first voice anti-spoofing collection in Spanish, covering five Latin-American accents (Argentina, Chile, Colombia, Peru, Venezuela).

Overview

The task is binary classification: bonafide (genuine human Spanish speech) vs. spoof (synthetic / converted speech). Bonafide audio is real Latin-American Spanish; spoof audio is produced by several generators:

  • CycleGAN, Diff(usion), StarGAN — voice-conversion across accent pairs (e.g. Argentina-Chile).
  • TTS — text-to-speech.
  • TTS-Diff, TTS-StarGAN — TTS output further passed through a Diff / StarGAN voice-conversion stage (the TTS-VC cascade).

License & redistribution

Redistributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license — the upstream HABLA license. See LICENSE.txt. CC BY 4.0 permits redistribution and derivative works with attribution. Labels and the evaluation protocol are unmodified; audio is shipped as the verbatim source 16 kHz mono PCM WAV (no re-encode — a whole-set decode probe confirmed every clip is already 16 kHz mono).

Schema

Column Type Description
path string source-relative path of the clip, unique
audio Audio(16000) 16 kHz mono
label ClassLabel "bonafide" (0) / "spoof" (1)
notes string JSON: utterance_id, speaker_id, system, source_subdir

utterance_id is the on-disk filename stem (globally unique across all six dirs). system is the spoof generator (CycleGAN / Diff / StarGAN / TTS / TTS-Diff / TTS-StarGAN) or real for bonafide.

notes example:

{"utterance_id": "CycleGAN-arf_00295_00101634363-arf_02121_002402", "speaker_id": "arf_00295", "system": "CycleGAN", "source_subdir": "CycleGAN/Argentina-Argentina/arf_00295-arf_02121"}

Quick Start 

from datasets import load_dataset

ds = load_dataset("SpeechAntiSpoofingBenchmarks/HABLA", split="test")
print(ds[0])

Stats

Stat Value
Total trials 80,816
Bonafide 22,816
Spoof 58,000
CycleGAN 16,000
Diff 16,000
StarGAN 16,000
TTS 5,000
TTS-Diff 2,500
TTS-StarGAN 2,500

Source provenance

Evaluation

For evaluation instructions and submission format, see submissions/README.md.

Citation 

@inproceedings{tamayoflorez23_interspeech,
  title     = {{HABLA: A Dataset of Latin American Spanish Accents for Voice Anti-spoofing}},
  author    = {Pablo Andrés {Tamayo Flórez} and Rubén Manrique and Bernardo {Pereira Nunes}},
  year      = {2023},
  booktitle = {{Interspeech 2023}},
  pages     = {1963--1967},
  doi       = {10.21437/Interspeech.2023-2272},
  issn      = {2958-1796},
}

Maintainer

Maintained by Kirill Borodin (SpeechAntiSpoofingBenchmarks).