Jabberjay 🦜

One API. Every state-of-the-art synthetic voice detector.

Why Jabberjay?

Synthetic voice detection is a fragmented landscape — state-of-the-art models are scattered across research repositories, each with its own dependencies, input formats, and output conventions. Jabberjay brings them all under one consistent Python API and CLI so you can detect AI-generated speech without wrestling with model internals.

• Ten model families — ViT, AST, Spectra0, SpectraAASIST, SpectraAASIST3, Wav2Vec2, HuBERT, WavLM, RawNet2, and a classical baseline
• Unified output — every model returns the same DetectionResult with label, confidence, and scores
• Zero boilerplate — pass a file path, get a verdict; models are downloaded and cached automatically
• Flexible — use strings for quick experiments, enums for IDE autocomplete, or pre-load audio to run multiple models on the same clip

$ jabberjay interview.wav
Bonafide ✔️  (94.1% confidence, model=VIT)

$ jabberjay suspicious.wav -m HuBERT
Spoof ❌  (97.8% confidence, model=HuBERT)

Installation

pip install jabberjay

Requires Python ≥ 3.12. Models are downloaded from Hugging Face Hub on first use and cached locally.

Quickstart

from Jabberjay import Jabberjay

jj = Jabberjay()
result = jj.detect("audio.wav")

print(result)              # Bonafide ✔️ (94.1% confidence, model=VIT)
print(result.label)        # "Bonafide"
print(result.is_bonafide)  # True
print(result.confidence)   # 0.941

Models

Vision Transformer (ViT)

Audio Spectrogram Transformer (AST)

Wav2Vec2

HuBERT

WavLM

Spectra Family (lab260)

Other

Usage

Command Line Interface

usage: jabberjay [-h] [-m {AST,Classical,HuBERT,RawNet2,Spectra0,SpectraAASIST,SpectraAASIST3,VIT,Wav2Vec2,WavLM}]
                 [-d {ASVspoof2019,ASVspoof5,VoxCelebSpoof}]
                 [-vis {ConstantQ,MelSpectrogram,MFCC}] [-v]
                 audio

# Quickstart — VIT with ConstantQ on VoxCelebSpoof (defaults)
jabberjay audio.wav

# Self-contained models (no dataset or visualisation required)
jabberjay audio.wav -m Spectra0
jabberjay audio.wav -m SpectraAASIST
jabberjay audio.wav -m SpectraAASIST3
jabberjay audio.wav -m Wav2Vec2
jabberjay audio.wav -m HuBERT
jabberjay audio.wav -m WavLM
jabberjay audio.wav -m RawNet2

# AST with a specific dataset
jabberjay audio.wav -m AST -d ASVspoof2019

# VIT with full options
jabberjay audio.wav -m VIT -d ASVspoof5 -vis MelSpectrogram

# Verbose output
jabberjay audio.wav -v

Python API

All public names are importable from the top-level package:

from Jabberjay import Jabberjay, DetectionResult, Model, Dataset, Visualisation

Choosing a model

String names and enum values are both accepted:

jj = Jabberjay()

# Self-contained models — no extra arguments needed
result = jj.detect("audio.wav", model="Spectra0")
result = jj.detect("audio.wav", model="SpectraAASIST")
result = jj.detect("audio.wav", model="SpectraAASIST3")
result = jj.detect("audio.wav", model="Wav2Vec2")
result = jj.detect("audio.wav", model="HuBERT")
result = jj.detect("audio.wav", model="WavLM")
result = jj.detect("audio.wav", model="RawNet2")
result = jj.detect("audio.wav", model="Classical")

# AST — requires a dataset
result = jj.detect("audio.wav", model="AST", dataset="VoxCelebSpoof")

# VIT — requires a dataset and a visualisation
result = jj.detect("audio.wav", model="VIT", dataset="ASVspoof5", visualisation="MFCC")

# Enums give IDE autocomplete and catch typos at import time
result = jj.detect(
    "audio.wav",
    model=Model.VIT,
    dataset=Dataset.ASVspoof5,
    visualisation=Visualisation.MFCC,
)

DetectionResult

Every call to detect() returns a DetectionResult regardless of the model used:

if result.is_bonafide:
    print(f"Genuine voice detected with {result.confidence:.1%} confidence")
else:
    print(f"Synthetic voice detected with {result.confidence:.1%} confidence")

# Full per-label scores
if result.scores:
    for entry in result.scores:
        print(f"  {entry['label']}: {entry['score']:.3f}")

Pre-loading audio

Use load() when running multiple models on the same clip to avoid re-reading the file. Each model is also cached in memory after its first detect() call, so calling it again later in the process reuses the already-loaded weights instead of reloading them:

audio = jj.load("audio.wav")  # returns (samples, sample_rate)

results = [
    jj.detect(audio, model="Wav2Vec2"),
    jj.detect(audio, model="HuBERT"),
    jj.detect(audio, model="VIT", dataset="VoxCelebSpoof", visualisation="ConstantQ"),
]

Discovering available options

jj.list_models()         # returns list[Model]
jj.list_datasets()       # returns list[Dataset]
jj.list_visualisations() # returns list[Visualisation]

Examples

The examples/ directory contains focused, runnable scripts:

just example                           # run quickstart
just run-example preloading_audio      # run a specific example
just run-all                           # full sweep (slow — downloads all models)

Developer Setup

Requires uv and just.

git clone https://github.com/MattyB95/Jabberjay.git
cd Jabberjay
just install   # install all dependencies including dev tools

See just --list for all available commands.

Contributing

Contributions are welcome — especially new models! See CONTRIBUTING.md for a full guide.

The quickest way to make an impact is to open a model request issue with a HuggingFace link and licence details.

Support

If Jabberjay saves you time, consider supporting its development:

• GitHub Sponsors — recurring or one-off support, directly on GitHub
• Ko-fi — buy me a coffee

Citation

If you use Jabberjay in your research, please cite it. GitHub's "Cite this repository" button (in the sidebar) will generate APA or BibTeX automatically from the CITATION.cff file, or you can use the entry below directly:

@software{boakes_jabberjay_2026,
  author  = {Boakes, Matthew},
  title   = {Jabberjay},
  year    = {2026},
  url     = {https://github.com/MattyB95/Jabberjay},
  version = {0.1.0},
  doi     = {10.5281/zenodo.19056977},
}

Archived versions are available on Zenodo. The concept DOI (10.5281/zenodo.19056977) always resolves to the latest release.

Acknowledgement

This work was supported, in whole or in part, by the Bill & Melinda Gates Foundation [INV-001309].