SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2111.02813 · NeurIPS · 2021

WaveFake: A Data Set to Facilitate Audio Deepfake Detection

Joel Frank, Lea Schönherr

arXiv · PDF · Open in the Atlas

Code that ran

6 of 7 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 6 have no contract verdict; 0 failures of our instrument; 1 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2111.02813.json · badge: /api/ran/2111.02813.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
rub-syssec/wavefake canonical (classifier label)
classifier basis not recorded on the link
6 ran · 1 no recorded run
FunctionStatusWhere it lives
blur_image Ran
ran on a synthesized input
rub-syssec/wavefake/dfadetect/attribution.py
code served (permissive licence) · get_code("c2b0efa076bf5fbe")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned ndarray, ndim 3, dtype family float, shape (3, 4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from scipy import ndimage

torch.manual_seed(0)

# Create a small spectrogram tensor (2D: height x width)
spectrogram = torch.randn(4, 3)

# Create a list of sigma values
sigmas = [0.5, 1.0, 1.5]

ARGS = {
    "spectrogram": spectrogram,
    "sigmas": sigmas,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 3, dtype family float, shape (3, 4, 3) — shape (3, 4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
feature_kwargs Ran
ran on a synthesized input
rub-syssec/wavefake/experiment_config.py
code served (permissive licence) · get_code("7d6f13721ee12d25")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    'lfcc': True
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
find_wav_files Ran
ran on a synthesized input
rub-syssec/wavefake/dfadetect/utils.py
code served (permissive licence) · get_code("230e65e3fcbee0aa")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path

# Create a temporary directory with some .wav files
temp_dir = tempfile.TemporaryDirectory()
temp_path = Path(temp_dir.name)

# Create some test .wav files in the directory and subdirectories
(temp_path / "audio1.wav").touch()
(temp_path / "audio2.wav").touch()

subdir = temp_path / "subdir"
subdir.mkdir()
(subdir / "audio3.wav").touch()

# Define ARGS
ARGS = {
    "path_to_dir": temp_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
integral_approximation Ran
ran on a synthesized input
rub-syssec/wavefake/dfadetect/attribution.py
code served (permissive licence) · get_code("9112ab69d2ec3a50")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (3, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# The function expects a tensor of gradients with shape (steps, *feature_dims)
# It will:
# 1. Index with [:-1] and [1:] to get consecutive pairs
# 2. Average those pairs
# 3. Take mean along axis 0
# So we need at least 2 steps in the first dimension

# Create a small gradient tensor: shape (5, 3, 4) representing 5 steps with 3x4 features
gradients = torch.randn(5, 3, 4)

ARGS = {
    "gradients": gradients
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (3, 4) — shape (3, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
score Ran
ran on a synthesized input
rub-syssec/wavefake/dfadetect/models/gaussian_mixture_model.py
code served (permissive licence) · get_code("371c6c7eedd45605")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.distributions as D
from torch import nn

torch.manual_seed(0)

class SimpleGMM(nn.Module):
    """Minimal GMM implementation for testing."""
    
    def __init__(self, k: int, data: torch.Tensor, covariance_type: str = 'full'):
        super().__init__()
        self.k = k
        self.covariance_type = covariance_type
        self._fitted = False
        self._eps = torch.finfo(torch.float32).eps
        
        d = data.size(1)
        
        # Initialize parameters
        self.logits = nn.Parameter(torch.zeros(k))
        self.loc = nn.Parameter(torch.randn(k, d))
        
        if covariance_type == 'full':
            self.cov = nn.Parameter(torch.stack([torch.eye(d) for _ in range(k)]))
        elif covariance_type == 'diag':
            self.cov = nn.Parameter(torch.ones(k, d))
        
        self._build_distributions()
        self._fitted = True
    
    def _build_distributions(self):
        self._mix = D.Categorical(logits=self.logits)
        if self.covariance_type == 'full':
            self._comp = D.MultivariateNormal(self.loc, self.cov)
        else:
            self._comp = D.Normal(self.loc, self.cov)
    
    def forward(self, X):
        """Compute weighted log probabilities."""
        if not self._fitted:
            raise RuntimeError('Model not fitted')
        
        # X shape: (n_samples, n_features)
        # Unsqueeze to (n_samples, 1, n_features) for broadcasting
        X_unsqueezed = X.unsqueeze(1)
        
        # Compute log probs and add mixture logits
        weighted_log_prob = self._comp.log_prob(X_unsqueezed) + torch.log_softmax(self._mix.logits, dim=-1)
        
        # Sum over components
        return torch.logsumexp(weighted_log_prob, dim=1, keepdim=True)


# Create test data
data_real = torch.randn(10, 4)
data_fake = torch.randn(10, 4)
test_data = torch.randn(2, 4)

# Create models
real_model = SimpleGMM(k=3, data=data_real, covariance_type='full')
fake_model = SimpleGMM(k=3, data=data_fake, covariance_type='full')

ARGS = {
    'real_model': real_model,
    'fake_model': fake_model,
    'data': test_data,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
score_batch Ran
ran on a synthesized input
rub-syssec/wavefake/dfadetect/models/gaussian_mixture_model.py
code served (permissive licence) · get_code("9edd332fbe6d1e73")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 3, dtype family float, shape (2, 5, 1) as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.distributions as D

torch.manual_seed(0)

# Create a concrete implementation of GMMBase for testing
class ConcreteGMM(nn.Module):
    def __init__(self, k: int, data: torch.Tensor, covariance_type: str = 'full') -> None:
        super().__init__()
        self.k = k
        self.covariance_type = covariance_type
        self._mix = None
        self._comp = None
        self._eps = torch.finfo(torch.float32).eps
        self._fitted = False
        self._initialize(data)
    
    def _initialize(self, data: torch.Tensor):
        d = data.size(1)
        pi = torch.full(fill_value=1.0 / self.k, size=[self.k])
        prob = torch.ones(len(data)) / len(data)
        loc = data[torch.multinomial(prob, num_samples=self.k)]
        
        if self.covariance_type == 'full':
            cov = torch.stack([torch.eye(d) for _ in range(self.k)])
        elif self.covariance_type == 'diag':
            cov = torch.stack([torch.ones(d) for _ in range(self.k)])
        
        self._initialize_parameters(pi, loc, cov)
        self._build_distributions()
    
    def _initialize_parameters(self, pi: torch.Tensor, loc: torch.Tensor, cov: torch.Tensor) -> None:
        self.register_buffer('logits', torch.log(pi))
        self.register_buffer('loc', loc)
        self.register_buffer('cov', cov)
    
    def _build_distributions(self):
        self._mix = D.Categorical(logits=self.logits)
        if self.covariance_type == 'full':
            self._comp = D.MultivariateNormal(self.loc, self.cov)
        else:
            self._comp = D.Normal(self.loc, torch.sqrt(self.cov))
        self._fitted = True
    
    def forward(self, X):
        if not self._fitted:
            raise RuntimeError('Model not fitted')
        
        if self.covariance_type == 'full':
            weighted_log_prob = self._comp.log_prob(X.unsqueeze(1)) + torch.log_softmax(self.logits, dim=-1)
        else:
            weighted_log_prob = self._comp.log_prob(X.unsqueeze(1)).sum(dim=-1) + torch.log_softmax(self.logits, dim=-1)
        
        return torch.logsumexp(weighted_log_prob, dim=1, keepdim=True)

# Create training data for initialization
train_data_real = torch.randn(10, 4)
train_data_fake = torch.randn(10, 4)

# Create real and fake models
real_model = ConcreteGMM(k=3, data=train_data_real, covariance_type='full')
fake_model = ConcreteGMM(k=3, data=train_data_fake, covariance_type='full')

# Create test data: shape (batch_dim=2, n_points=5, feat_d=4)
test_data = torch.randn(2, 5, 4)

ARGS = {
    'real_model': real_model,
    'fake_model': fake_model,
    'data': test_data,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (2, 5, 1) — shape (2, 5, 1) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_gradients Not yet run
unverified (harvested, no recorded run)
rub-syssec/wavefake/dfadetect/attribution.py
code served (permissive licence) · get_code("055ad46b81aadbc9")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 7 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Deep generative modeling has the potential to cause significant harm to society. Recognizing this threat, a magnitude of research into detecting so-called "Deepfakes" has emerged. This research most often focuses on the image domain, while studies exploring generated audio signals have-so far -been neglected. In this paper, we aim to narrow this gap. We present a novel data set, for which we collected ten sample sets from six different network architectures, spanning two languages. We analyze the frequency statistics comprehensively, discovering subtle differences between the architectures, specifically among the higher frequencies. Additionally, to facilitate further development of detection methods, we implemented three different classifiers adopted from the signal processing community to give practitioners a baseline to compare against. In a first evaluation, we already discovered significant trade-offs between the different approaches. Neural networkbased approaches performed better on average, but more traditional models proved to be more robust.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2111.02813")
get_code_for_paper("2111.02813")
have("2111.02813")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2111.02813.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2111.02813.svg)](https://syntology.ai/paper/2111.02813)

Connect an agent — have() is free.