Embedding methods

The descriptions below state what the packaged code does. References identify the underlying technique or related study; they do not certify an exact reproduction of the paper’s implementation, training procedure or reported robustness. The review 1 provides background for the classical families, rather than a primary derivation of every local variant.

All methods accept a waveform and binary payload and expose blind extraction with the payload length. Use registry identifiers in experiment configurations. Compare capacity, BER, audio distortion and detectability separately under the experimental protocol.

30 methods are registered. Select a name for its mechanism, parameters and limits.

Method / registry identifier	What it does	Family · purpose	Reference
FBS-LSB FBSMethod	Adapts LSB capacity to foreground and background regions of the recording.	Sample-domain LSB · steganography	23
LSB LSB_METHOD	Hides one message bit in each audio sample with a very small numerical change.	Sample-domain LSB · steganography	1
Prime-factor interpolation PRIME_FACTOR_INTERPOLATE	Stores a variable number of bits as small offsets from interpolated sample values.	Sample-domain LSB · steganography	21
Blind SVD BLIND_SVD_METHOD	Embeds one bit per frame in the largest singular value of an entropy-selected DCT sub-band.	Transform domain · watermarking	20
DCT-b1 DCT_B1_METHOD	Embeds a watermark in a selected DCT band using masking and energy compensation.	Transform domain · watermarking	2
DCT-Delta-LSB DCT_DELTA_LSB_METHOD	Encodes bits in the relative norms of two groups of cosine-transform coefficients.	Transform domain · steganography	1
DWT-LSB DWT_LSB_METHOD	Hides bits in the parity of quantised wavelet detail coefficients.	Transform domain · steganography	1
EMD EMD_METHOD	Encodes bits in extrema of a slow component obtained by empirical mode decomposition.	Transform domain · watermarking	48
FSVC FSVC_METHOD	Stores bits in the relative singular values of two frequency-domain frame halves.	Transform domain · watermarking	5
LWT LWT_METHOD	Encodes bits as the sign of blocks of wavelet detail coefficients.	Transform domain · watermarking	22
Norm-space NORM_SPACE_METHOD	Represents each bit by which of two transform-domain vectors has the larger norm.	Transform domain · watermarking	4
Sync-DWT-DCT SYNC_DWT_DCT_METHOD	Combines transform-domain data with synchronisation codes to locate marked blocks after timing changes.	Transform domain · watermarking	47
Wireless DWT-LSB WIRELESS_DWT_LSB_METHOD	Places payload bits in the low-frequency approximation coefficients of a wavelet transform.	Transform domain · steganography	27
DSSS DSSS_METHOD	Spreads each bit over a frame using a key-derived pseudo-noise sequence.	Spread spectrum · steganography	6
ISS IMPROVED_SPREAD_SPECTRUM_METHOD	Reduces host-signal interference before adding a spread-spectrum watermark.	Spread spectrum · watermarking	30
Backward–forward echo BACKWARD_FORWARD_ECHO_METHOD	Carries bits with paired backward and forward echoes.	Echo hiding · watermarking	32
Echo hiding ECHO_METHOD	Represents zero and one by echoes with different delays.	Echo hiding · steganography	1
Time-spread echo TIME_SPREAD_ECHO_METHOD	Distributes the echo watermark over multiple delays using a keyed sequence.	Echo hiding · watermarking	33
Improved phase coding IMPROVED_PHASE_CODING_METHOD	Distributes phase-coded message portions across multiple Fourier blocks.	Phase coding · steganography	19
Phase coding PHASE_CODING_METHOD	Stores the message in Fourier phases of the first audio block.	Phase coding · steganography	1
QIM / ST-DM QIM_METHOD	Encodes bits by choosing between two quantisation grids for a keyed frame projection.	Quantisation (QIM) · watermarking	29
Histogram HISTOGRAM_METHOD	Stores bits in the distribution of sample amplitudes rather than their positions.	Statistical / patchwork · watermarking	34
LFAM LOW_FREQUENCY_AMPLITUDE_METHOD	Encodes bits through amplitude relations between three low-frequency sub-segments.	Statistical / patchwork · watermarking	35
Patchwork-ML PATCHWORK_MULTILAYER_METHOD	Carries bits in statistical differences between paired DCT coefficient groups.	Statistical / patchwork · watermarking	3
AAC-STC AAC_STC_METHOD	Chooses low-cost sample edits using perceptual codec residuals and syndrome-trellis coding.	Adaptive coding · steganography	25
PEE REVERSIBLE_PEE_METHOD	Hides data while allowing exact recovery of the original 16-bit PCM cover when no damage occurs.	Reversible (lossless) · steganography	49, 50
LE-GA LEARNABLE_EMBEDDING_GA_METHOD	Provides an experimental embedding pipeline inspired by learnable watermarking and genetic optimisation.	Learned embedding · watermarking	28
AudioSeal AUDIOSEAL_METHOD	Uses pretrained AudioSeal networks to generate and detect a neural audio watermark.	Neural network · watermarking	36
FGAS FGAS_METHOD	Optimises a small audio perturbation so a fixed neural decoder outputs the desired bits.	Neural network · steganography	24
WavMark WAVMARK_METHOD	Uses the pretrained WavMark neural model to embed and recover bit payloads in audio chunks.	Neural network · watermarking	37

Interpretation

Steganographic security concerns detecting the presence of a payload; watermark robustness concerns recovering it after processing. High SNR alone establishes neither. Paper-reported performance must retain its corpus, payload rate, attack severity and detector assumptions when compared with local results.

The LSB and FBS variants work on floating-point representations. A successful in-memory round trip does not imply survival after PCM export. Patchwork-ML, LWT and LE-GA have explicit implementation restrictions in the table; especially LE-GA must not be reported as a reproduced trained neural system.

Sync-DWT-DCT is the only classical method that searches the signal for a synchronisation code; frame-based methods that read a fixed grid lose synchronisation when samples are removed or inserted at the start. It reads a message longer than one block (32 bits by default) by numbering blocks from the first one found, so cropping the start preserves only a message that fits in one block. EMD depends on a data-driven decomposition that processing itself changes: low-pass filtering in particular alters which content forms the carrier, so the method is markedly less robust to filtering than fixed-basis transforms. PEE is reversible, not robust: any change to the marked samples, including requantisation, destroys both the payload and the ability to restore the cover, and a cover that is not on the 16-bit grid is restored in its rounded form.

AudioSeal and WavMark require the-a-files[neural] and pretrained weights. FGAS requires the-a-files[ai]. See installation and provenance.

Method reference

Generated from the card attribute of each class by python -m taf.catalogue_docs; edit the card, not this section. The same texts, in English and Polish, appear in the research UI.

Sample-domain LSB

FBS-LSB

FBSMethod · Sample-domain LSB · steganography · source code

Adapts LSB capacity to foreground and background regions of the recording.

Separates foreground from background, shuffles sample positions with a shared seed, then stores two float-representation bits per foreground sample and one per background sample. Decoding repeats the segmentation and ordering. Both sample precision and stable segmentation matter; format conversion or processing can break extraction.

Parameter	Default	Role
seed	42	secret key

Reference: Wang & Wang (2025) [23]

LSB

LSB_METHOD · Sample-domain LSB · steganography · source code

Hides one message bit in each audio sample with a very small numerical change.

Replaces the least significant bit of the sample’s 32-bit floating-point representation. Decoding reads those bits in order. This implementation uses float bits, not integer PCM bits: conversion to PCM, lossy compression or even small signal processing changes can erase the message.

No tunable parameters.

Reference: Alsabhany et al. (2020) [1]

Prime-factor interpolation

PRIME_FACTOR_INTERPOLATE · Sample-domain LSB · steganography · source code

Stores a variable number of bits as small offsets from interpolated sample values.

Every second sample is replaced by an interpolation of its retained neighbours plus the payload value. Capacity follows the least prime factor of a log-scaled neighbour difference, capped by max_bits_per_sample. The decoder derives the same prediction and capacity from the neighbours. More bits mean larger offsets; changing the neighbours can corrupt decoding. Recovering the original cover is not part of the TAF interface, so reversibility is not established here.

Parameter	Default	Role
max_bits_per_sample	4

Reference: Adhiyaksa et al. (2022) [21]

Transform domain

Blind SVD

BLIND_SVD_METHOD · Transform domain · watermarking · source code

Embeds one bit per frame in the largest singular value of an entropy-selected DCT sub-band.

Chooses a low-frequency sub-band with maximum estimated entropy, reshapes it into a matrix and applies SVD. Quantisation parity of the largest singular value carries the bit. The step is derived from the remaining singular values and quantization_coefficient. No original cover is needed, but processing that changes sub-band selection can disrupt detection.

Parameter	Default	Role
frame_size	1024
sub_band_count	4
quantization_coefficient	0.1	embedding strength

Reference: Dhar & Shimamura (2015) [20]

DCT-b1

DCT_B1_METHOD · Transform domain · watermarking · source code

Embeds a watermark in a selected DCT band using masking and energy compensation.

Processes audio in frames, changes coefficient groups to represent bits, and compensates energy changes while reconstructing the waveform. The detector evaluates the corresponding coefficient relations. Frame and group lengths control capacity and the embedding region; preserving frame alignment is essential after temporal attacks.

Parameter	Default	Role
lt	23
lw	1486
lG1	24
lG2	6
key	20240521	secret key

Reference: Hu & Hsu (2015) [2]

DCT-Delta-LSB

DCT_DELTA_LSB_METHOD · Transform domain · steganography · source code

Encodes bits in the relative norms of two groups of cosine-transform coefficients.

Transforms each frame with DCT, splits coefficients into two vectors and makes one norm larger than the other according to the bit. Decoding compares the norms. Despite the historical LSB name, this implementation modifies norm relations. delta_value sets the separation relative to the frame level; larger separation increases distortion and the decision margin.

Parameter	Default	Role
frame_length_in_ms	100
delta_value	0.05	embedding strength

Reference: Alsabhany et al. (2020) [1]

DWT-LSB

DWT_LSB_METHOD · Transform domain · steganography · source code

Hides bits in the parity of quantised wavelet detail coefficients.

A two-level discrete wavelet transform separates the signal into approximation and detail bands. Selected detail coefficients are quantised to even or odd indices and decoded by parity. step_scale controls the step relative to detail-band RMS, making it follow signal level. Filtering and requantisation can still move coefficients across decision boundaries.

Parameter	Default	Role
dwt_type	'bior5.5'
step_scale	0.5	embedding strength
spacing	10

Reference: Alsabhany et al. (2020) [1]

EMD

EMD_METHOD · Transform domain · watermarking · source code

Encodes bits in extrema of a slow component obtained by empirical mode decomposition.

Decomposes frames into oscillatory modes and quantises extrema of the remainder after the selected IMFs. It re-decomposes and corrects embedding in a closed loop; extrema vote on one bit per frame. This differs from marking only the last IMF in the paper. Filtering can change the decomposition itself, while repeated decomposition also increases computation cost. The paper's synchronisation code is not reproduced; Sync-DWT-DCT provides a sync-code scheme.

Parameter	Default	Role
frame_length	1024
imf_count	4
step_scale	0.2	embedding strength
sift_iterations	4
edge_margin	64
max_refinements	6

Reference: Khaldi & Boudraa (2013) [48]

FSVC

FSVC_METHOD · Transform domain · watermarking · source code

Stores bits in the relative singular values of two frequency-domain frame halves.

Splits each payload frame in half, applies DCT, selects coefficient bands and computes their singular values. Embedding enforces a bit-dependent ratio; decoding compares the values. In the current implementation the ratio is controlled by an internal alpha; the exposed delta parameter is stored but does not affect encoding. Temporal misalignment changes the compared halves.

Parameter	Default	Role
delta	3.5793656	embedding strength

Reference: Zhao et al. (2021) [5]

LWT

LWT_METHOD · Transform domain · watermarking · source code

Encodes bits as the sign of blocks of wavelet detail coefficients.

This implementation uses a two-level Haar wavelet decomposition and groups eight detail coefficients per bit. It forces a positive or negative block with a minimum magnitude set by threshold, then reconstructs the signal. Decoding reads the sign of the block mean. A larger threshold creates a stronger mark and more distortion; alignment and detail-band preservation matter. The transform is the PyWavelets Haar decomposition rather than an explicit lifting scheme.

Parameter	Default	Role
threshold	0.05	embedding strength

Reference: Mushtaq et al. (2024) [22]

Norm-space

NORM_SPACE_METHOD · Transform domain · watermarking · source code

Represents each bit by which of two transform-domain vectors has the larger norm.

Applies a Haar DWT and then DCT to the approximation band. Even and odd DCT coefficients form two vectors whose norms are separated by a relative delta. The decoder compares them without the original recording. Increasing delta strengthens the separation but changes audio more; cropping can break the segment grid.

Parameter	Default	Role
delta	0.05	embedding strength

Reference: Saadi et al. (2019) [4]

Sync-DWT-DCT

SYNC_DWT_DCT_METHOD · Transform domain · watermarking · source code

Combines transform-domain data with synchronisation codes to locate marked blocks after timing changes.

Writes a Barker-based sync pattern through quantised sample-group means, followed by payload quantisation in the DCT of a DWT approximation band. The decoder scans offsets and validates candidate blocks. Relative quantisation steps follow signal level. Coefficient selection and step rules are implementation choices; successful resynchronisation still depends on the attack and remaining audio. The sync code has 16 bits (the 13-bit Barker code followed by a 3-bit one), one bit per sync_group samples; 16-sample groups replace the paper's 5. Because the decoder searches every offset, it finds the blocks again after shifting, padding or cropping.

Parameter	Default	Role
segment_length	4096
sync_group	16
level	3
wavelet	'db4'
band_start	200.0
bits_per_block	32
step_scale	0.5	embedding strength
sync_scale	0.3
detection_threshold	0.2

Reference: Wang & Zhao (2006) [47]

Wireless DWT-LSB

WIRELESS_DWT_LSB_METHOD · Transform domain · steganography · source code

Places payload bits in the low-frequency approximation coefficients of a wavelet transform.

Scales and rounds DWT approximation coefficients to integers, replaces up to lsb_depth low bits and reconstructs the waveform. Decoding repeats the transform and scaling. Wavelet type, level and coefficient_scale affect capacity and distortion. This adapts the paper’s carrier idea to direct bit payloads; the name does not imply guaranteed survival of a wireless channel.

Parameter	Default	Role
dwt_type	'haar'
level	1
lsb_depth	8
coefficient_scale	22000

Reference: Hamdi et al. (2025) [27]

Spread spectrum

DSSS

DSSS_METHOD · Spread spectrum · steganography · source code

Spreads each bit over a frame using a key-derived pseudo-noise sequence.

Adds a bipolar sequence with sign determined by the bit and amplitude proportional to local RMS and alpha. The receiver correlates the frame with the same sequence and reads the correlation sign. Spreading distributes the payload over many samples, but the cover itself interferes with detection. Larger alpha improves the detection margin at the cost of more added noise.

Parameter	Default	Role
key	20240521	secret key
alpha	0.05	embedding strength
min_chip_length	1024

Reference: Nugraha (2011) [6] · Needs long input: short files produce failed trials.

ISS

IMPROVED_SPREAD_SPECTRUM_METHOD · Spread spectrum · watermarking · source code

Reduces host-signal interference before adding a spread-spectrum watermark.

Projects each carrier frame onto a key-derived chip sequence and compensates its existing component before setting a bit-dependent projection. The decoder reads the correlation sign. This removes a source of ambiguity present in plain DSSS. strength controls the mark level; compensation itself also changes the carrier, and frame synchronisation is still required.

Parameter	Default	Role
key	20240521	secret key
strength	0.05	embedding strength
removal	1.0
min_chip_length	256

Reference: Malvar & Florencio (2003) [30]

Echo hiding

Backward–forward echo

BACKWARD_FORWARD_ECHO_METHOD · Echo hiding · watermarking · source code

Carries bits with paired backward and forward echoes.

Uses echo kernels on both sides of a sample position with opposite signs. The detector compares the corresponding cepstral responses rather than relying on a single echo peak. alpha controls the added echo strength. The paired structure changes the detection statistic, but does not remove sensitivity to reverberation or lost frame alignment.

Parameter	Default	Role
alpha	0.1	embedding strength
d0	150
d1	200
min_frame_length	2048

Reference: Kim & Choi (2003) [32]

Echo hiding

ECHO_METHOD · Echo hiding · steganography · source code

Represents zero and one by echoes with different delays.

Splits the recording into payload frames and mixes in a delayed copy using alpha. The decoder detects the selected delay through cepstral analysis. Echo strength trades audibility against detection margin. Short frames, existing reverberation or temporal changes can make the two delay hypotheses difficult to distinguish.

Parameter	Default	Role
alpha	0.2	embedding strength
d0	150
d1	200
min_frame_length	2048

Reference: Alsabhany et al. (2020) [1] · Needs long input: short files produce failed trials.

Time-spread echo

TIME_SPREAD_ECHO_METHOD · Echo hiding · watermarking · source code

Distributes the echo watermark over multiple delays using a keyed sequence.

A pseudo-noise kernel spreads echo energy in time instead of concentrating it at one delay. The decoder uses the corresponding sequence to detect the bit from the cepstral response. alpha sets strength and the shared seed reproduces the kernel. Spreading changes the audibility and detection trade-off; it still needs enough audio per bit and correct alignment.

Parameter	Default	Role
key	20240521	secret key
alpha	0.1	embedding strength
pn_length	511
d0	200
d1	900
min_frame_length	4096

Reference: Ko et al. (2005) [33]

Phase coding

Improved phase coding

IMPROVED_PHASE_CODING_METHOD · Phase coding · steganography · source code

Distributes phase-coded message portions across multiple Fourier blocks.

Divides the payload among segments and sets selected phase pairs to ±π/2 while retaining spectral magnitudes. The decoder reconstructs the segment layout from the message length and reads phase signs. Distribution avoids concentrating all bits in the first segment, but it is not error correction; altered length, phase distortion and trimming can still corrupt bits.

No tunable parameters.

Reference: Yang (2024) [19]

Phase coding

PHASE_CODING_METHOD · Phase coding · steganography · source code

Stores the message in Fourier phases of the first audio block.

Sets selected phases to +π/2 or −π/2 and mirrors them to preserve a real waveform. Subsequent blocks retain their original inter-block phase differences. Decoding reads the signs of the first block’s phases. Magnitudes are retained during embedding, but a shift or removal of the opening block can destroy access to the payload.

No tunable parameters.

Reference: Alsabhany et al. (2020) [1]

Quantisation (QIM)

QIM / ST-DM

QIM_METHOD · Quantisation (QIM) · watermarking · source code

Encodes bits by choosing between two quantisation grids for a keyed frame projection.

Projects a frame onto a key-derived direction and moves that projection to the nearest point on the bit’s dithered grid. The decoder chooses the closer grid. The step follows energy orthogonal to the projection, which embedding leaves unchanged, and is scaled by step_scale. Larger steps increase separation and distortion; timing changes disrupt the frame projections.

Parameter	Default	Role
key	20240521	secret key
step_scale	0.1	embedding strength
min_frame_length	256

Reference: Chen & Wornell (2001) [29]

Statistical / patchwork

Histogram

HISTOGRAM_METHOD · Statistical / patchwork · watermarking · source code

Stores bits in the distribution of sample amplitudes rather than their positions.

Forms an amplitude histogram and modifies population relations in groups of three neighbouring bins. The histogram range is relative to mean absolute amplitude; threshold controls the required relation. Because sample order is not used, the method targets timing and cropping robustness. Changes that reshape amplitude statistics, such as clipping or noise, can still destroy the relation.

Parameter	Default	Role
amplitude_span	2.5
threshold	2.0	embedding strength
cutoff	2000.0
rounds	24
search_span	0.2
search_steps	41
min_samples_per_bin	32

Reference: Xiang & Huang (2007) [34]

LFAM

LOW_FREQUENCY_AMPLITUDE_METHOD · Statistical / patchwork · watermarking · source code

Encodes bits through amplitude relations between three low-frequency sub-segments.

Separates low-frequency content and scales consecutive sub-segments so the middle amplitude lies above or below the average of its neighbours. The decoder reads that relation without the original. margin controls the separation. Relative amplitudes tolerate common gain changes, but high-pass filtering and temporal misalignment can remove or rearrange the carrier.

Parameter	Default	Role
cutoff	2000.0
margin	0.3	embedding strength
min_segment_length	768

Reference: Lie & Chang (2006) [35]

Patchwork-ML

PATCHWORK_MULTILAYER_METHOD · Statistical / patchwork · watermarking · source code

Carries bits in statistical differences between paired DCT coefficient groups.

Reorders coefficients in a selected band and adjusts the mean absolute values of paired segments so their ordering represents a bit. Extraction compares these means. Despite the multilayer name, the current code implements only the first layer. min_segment_length limits capacity; band removal and changes to the global transform grid can damage the mark.

Parameter	Default	Role
min_segment_length	8

Reference: Natgunanathan et al. (2017) [3]

Adaptive coding

AAC-STC

AAC_STC_METHOD · Adaptive coding · steganography · source code

Chooses low-cost sample edits using perceptual codec residuals and syndrome-trellis coding.

Converts audio to 16-bit PCM and estimates embedding costs from the difference after an AAC round trip, with Vorbis as a fallback. A trellis search selects inexpensive ±1 changes whose LSB syndrome encodes the message. Decoding applies the same parity-check matrix. The codec guides where to embed; it does not make the resulting LSB payload inherently resistant to transcoding.

Parameter	Default	Role
bitrate	400000
constraint_height	7
hhat_seed	0	secret key

Reference: Luo et al. (2017) [25]

Reversible (lossless)

PEE

REVERSIBLE_PEE_METHOD · Reversible (lossless) · steganography · source code

Hides data while allowing exact recovery of the original 16-bit PCM cover when no damage occurs.

Predicts each sample from its two predecessors and expands small prediction errors as 2e + bit; larger errors are shifted. A location map handles overflow risks and decoding reverses the mappings. threshold controls eligible errors and capacity. Reversibility refers to the cover rounded onto the 16-bit grid; processing or another bit-depth conversion can destroy both payload and restoration. recover_cover() returns the original cover exactly from an undamaged stego signal. The method is fragile by design.

Parameter	Default	Role
threshold	8	embedding strength

Reference: Thodi & Rodriguez (2007) [49], Nishimura (2011) [50]

Learned embedding

LE-GA

LEARNABLE_EMBEDDING_GA_METHOD · Learned embedding · watermarking · source code

Provides an experimental embedding pipeline inspired by learnable watermarking and genetic optimisation.

Maps bits to an upscaled waveform, applies a mask and adds a scaled residual; extraction uses correlation-style pooling. Without the paper’s trained weights, this implementation uses deterministic minimal operators and an all-ones mask. embedding_strength sets the residual level. Treat it as an implementation-specific baseline, not a reproduction of the trained network’s reported performance.

Parameter	Default	Role
scaling_factor	0.9
embedding_strength	0.005	embedding strength
population_size	10
mutation_rate	0.1
mutation_std	0.05
seed	42	secret key
min_samples_per_bit	16

Reference: Nayeem et al. (2026) [28]

Neural network

AudioSeal

AUDIOSEAL_METHOD · Neural network · watermarking · source code

Uses pretrained AudioSeal networks to generate and detect a neural audio watermark.

Wraps the optional AudioSeal package and released model weights. The generator carries 16 bits per chunk, so longer messages are split across consecutive chunks; alpha scales the watermark contribution. Detection uses the corresponding neural model. Dependencies and weights must be available, and robustness should be measured for the actual audio and attack settings.

Parameter	Default	Role
generator	'audioseal_wm_16bits'
detector	'audioseal_detector_16bits'
alpha	1.0	embedding strength
min_chunk_length	16000

Reference: San Roman et al. (2024) [36] · Requires: audioseal, torch — install with pip install "the-a-files[neural]"

FGAS

FGAS_METHOD · Neural network · steganography · source code

Optimises a small audio perturbation so a fixed neural decoder outputs the desired bits.

Builds a deterministically initialised 1D convolutional decoder from a shared seed and optimises the input perturbation against the target message. The receiver reconstructs that decoder and thresholds its output. epsilon bounds the perturbation; optimisation settings affect cost and success. A small perturbation does not by itself guarantee robustness to codecs or resistance to steganalysis.

Parameter	Default	Role
hidden_channels	16
key	1337	secret key
iterations	300
epsilon	0.02	embedding strength
learning_rate	0.005
alpha	50.0
beta	0.5
leaky_slope	0.2
radius	16

Reference: Yan et al. (2025) [24] · Requires: tensorflow — install with pip install "the-a-files[ai]"

WavMark

WAVMARK_METHOD · Neural network · watermarking · source code

Uses the pretrained WavMark neural model to embed and recover bit payloads in audio chunks.

A one-second model window carries 32 bits: 16 synchronisation bits and 16 payload bits. The decoder searches for the sync pattern; this adapter uses 16 kHz audio and defaults to chunks of at least two seconds to leave search space. Longer messages use consecutive chunks. The package and weights are optional dependencies; robustness depends on the model and evaluated channel.

Parameter	Default	Role
min_chunk_length	32000

Reference: Chen et al. (2023) [37] · Requires: wavmark, torch — install with pip install "the-a-files[neural]"