SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.26515 · 2026

JAL-Turn: Joint Acoustic-Linguistic Modeling for Real-Time and Robust Turn-Taking Detection in Full-Duplex Spoken Dialogue Systems

Yu Pan, Shi Qiu, Guangzhao Yang, Ningjie Bai

arXiv · PDF · Open in the Atlas

Code that ran

17 of 32 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, 2 violated it and 10 have no contract verdict; 5 failures of our instrument; 15 no recorded run), measured 2026-11-22; 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/2603.26515.json · badge: /api/ran/2603.26515.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
pipecat-ai/smart-turn canonical (classifier label)
classifier basis: census
3 ran · 1 no recorded run
inokoj/VAP-Realtime community
Papers with Code archive
13 ran · 14 no recorded run
ten-framework/ten-turn-detection community
Papers with Code archive
1 ran · 0 no recorded run
FunctionStatusWhere it lives
Base Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("676ee8e8bc273549")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned NoneType. 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
import numpy as np

INIT_ARGS = {}

ARGS = {}

METHOD = 'start_process'
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type NoneType — a NoneType output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a NoneType 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
CPCAR Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("fd6ea23082a35392")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, 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 torch
import torch.nn as nn

class CPCAR(nn.Module):
    """
    Most of the code in this file are scaled down (and heavily copied) versions of
        https://github.com/facebookresearch/CPC_audio
    """

    def __init__(self, dimEncoded, dimOutput, keepHidden, nLevelsGRU, mode='GRU', reverse=False):
        super(CPCAR, self).__init__()
        self.RESIDUAL_STD = 0.1
        if mode == 'LSTM':
            self.baseNet = nn.LSTM(dimEncoded, dimOutput, num_layers=nLevelsGRU, batch_first=True)
        elif mode == 'RNN':
            self.baseNet = nn.RNN(dimEncoded, dimOutput, num_layers=nLevelsGRU, batch_first=True)
        else:
            self.baseNet = nn.GRU(dimEncoded, dimOutput, num_layers=nLevelsGRU, batch_first=True)
        self.hidden = None
        self.keepHidden = keepHidden
        self.reverse = reverse

    def getDimOutput(self):
        return self.baseNet.hidden_size

    def forward(self, x):
        if self.reverse:
            x = torch.flip(x, [1])
        try:
            self.baseNet.flatten_parameters()
        except RuntimeError:
            pass
        x, h = self.baseNet(x, self.hidden)
        if self.keepHidden:
            if isinstance(h, tuple):
                self.hidden = tuple((x.detach() for x in h))
            else:
                self.hidden = h.detach()
        if self.reverse:
            x = torch.flip(x, [1])
        return x


INIT_ARGS = {
    'dimEncoded': 4,
    'dimOutput': 8,
    'keepHidden': True,
    'nLevelsGRU': 2,
    'mode': 'GRU',
    'reverse': False,
}

ARGS = {
    'x': torch.randn(2, 5, 4),
}

if __name__ == '__main__':
    model = CPCAR(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output dtype: {output.dtype}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
CPCEncoder Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("579b6ca8c9508891")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, 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 torch
import torch.nn as nn
import torch.nn.functional as F

class ChannelNorm(nn.Module):
    """
    Most of the code in this file are scaled down (and heavily copied) versions of
        https://github.com/facebookresearch/CPC_audio
    """

    def __init__(self, numFeatures, epsilon=1e-05, affine=True):
        super(ChannelNorm, self).__init__()
        if affine:
            self.weight = nn.parameter.Parameter(torch.Tensor(1, numFeatures, 1))
            self.bias = nn.parameter.Parameter(torch.Tensor(1, numFeatures, 1))
        else:
            self.weight = None
            self.bias = None
        self.epsilon = epsilon
        self.p = 0
        self.affine = affine
        self.reset_parameters()

    def reset_parameters(self):
        if self.affine:
            torch.nn.init.ones_(self.weight)
            torch.nn.init.zeros_(self.bias)

    def forward(self, x):
        cumMean = x.mean(dim=1, keepdim=True)
        cumVar = x.var(dim=1, keepdim=True)
        x = (x - cumMean) * torch.rsqrt(cumVar + self.epsilon)
        if self.weight is not None:
            x = x * self.weight + self.bias
        return x

class CPCEncoder(nn.Module):
    """
    Most of the code in this file are scaled down (and heavily copied) versions of
        https://github.com/facebookresearch/CPC_audio
    """

    def __init__(self, sizeHidden=512, normMode='layerNorm'):
        super(CPCEncoder, self).__init__()
        normLayer = ChannelNorm
        self.dimEncoded = sizeHidden
        self.conv0 = nn.Conv1d(1, sizeHidden, 10, stride=5, padding=3)
        self.batchNorm0 = normLayer(sizeHidden)
        self.conv1 = nn.Conv1d(sizeHidden, sizeHidden, 8, stride=4, padding=2)
        self.batchNorm1 = normLayer(sizeHidden)
        self.conv2 = nn.Conv1d(sizeHidden, sizeHidden, 4, stride=2, padding=1)
        self.batchNorm2 = normLayer(sizeHidden)
        self.conv3 = nn.Conv1d(sizeHidden, sizeHidden, 4, stride=2, padding=1)
        self.batchNorm3 = normLayer(sizeHidden)
        self.conv4 = nn.Conv1d(sizeHidden, sizeHidden, 4, stride=2, padding=1)
        self.batchNorm4 = normLayer(sizeHidden)
        self.DOWNSAMPLING = 160

    def getDimOutput(self):
        return self.conv4.out_channels

    def forward(self, x):
        x = F.relu(self.batchNorm0(self.conv0(x)))
        x = F.relu(self.batchNorm1(self.conv1(x)))
        x = F.relu(self.batchNorm2(self.conv2(x)))
        x = F.relu(self.batchNorm3(self.conv3(x)))
        x = F.relu(self.batchNorm4(self.conv4(x)))
        return x


INIT_ARGS = {
    'sizeHidden': 64,
    'normMode': 'layerNorm'
}

ARGS = {
    'x': torch.randn(2, 1, 16000)
}

if __name__ == '__main__':
    model = CPCEncoder(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Model output dim: {model.getDimOutput()}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
ChannelNorm Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("b4d2cb5bfcf2d4f4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float, shape (2, 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 torch.nn as nn

class ChannelNorm(nn.Module):
    """
    Most of the code in this file are scaled down (and heavily copied) versions of
        https://github.com/facebookresearch/CPC_audio
    """

    def __init__(self, numFeatures, epsilon=1e-05, affine=True):
        super(ChannelNorm, self).__init__()
        if affine:
            self.weight = nn.parameter.Parameter(torch.Tensor(1, numFeatures, 1))
            self.bias = nn.parameter.Parameter(torch.Tensor(1, numFeatures, 1))
        else:
            self.weight = None
            self.bias = None
        self.epsilon = epsilon
        self.p = 0
        self.affine = affine
        self.reset_parameters()

    def reset_parameters(self):
        if self.affine:
            torch.nn.init.ones_(self.weight)
            torch.nn.init.zeros_(self.bias)

    def forward(self, x):
        cumMean = x.mean(dim=1, keepdim=True)
        cumVar = x.var(dim=1, keepdim=True)
        x = (x - cumMean) * torch.rsqrt(cumVar + self.epsilon)
        if self.weight is not None:
            x = x * self.weight + self.bias
        return x


INIT_ARGS = {
    'numFeatures': 4,
    'epsilon': 1e-05,
    'affine': True,
}

ARGS = {
    'x': torch.randn(2, 4, 3),
}

if __name__ == '__main__':
    model = ChannelNorm(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output: {output}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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, 4, 3) — shape (2, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Codebook Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("ce669279019832c0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 1, dtype family int, shape (2,) 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 torch.nn as nn
from torch import Tensor
from einops import rearrange

class Codebook(nn.Module):

    def __init__(self, bin_frames):
        super().__init__()
        self.bin_frames = bin_frames
        self.n_bins: int = len(self.bin_frames)
        self.total_bins: int = self.n_bins * 2
        self.n_classes: int = 2 ** self.total_bins
        self.emb = nn.Embedding(num_embeddings=self.n_classes, embedding_dim=self.total_bins)
        self.emb.weight.data = self.create_code_vectors(self.total_bins)
        self.emb.weight.requires_grad_(False)

    def single_idx_to_onehot(self, idx: int, d: int=8) -> Tensor:
        assert idx < 2 ** d, 'must be possible with {d} binary digits'
        z = torch.zeros(d)
        b = bin(idx).replace('0b', '')
        for i, v in enumerate(b[::-1]):
            z[i] = float(v)
        return z

    def create_code_vectors(self, n_bins: int) -> Tensor:
        """
        Create a matrix of all one-hot encodings representing a binary sequence of `self.total_bins` places
        Useful for usage in `nn.Embedding` like module.
        """
        n_codes = 2 ** n_bins
        embs = torch.zeros((n_codes, n_bins))
        for i in range(2 ** n_bins):
            embs[i] = self.single_idx_to_onehot(i, d=n_bins)
        return embs

    def encode(self, x: Tensor) -> Tensor:
        """

        Encodes projection_windows x (*, 2, 4) to indices in codebook (..., 1)

        Arguments:
            x:          Tensor (*, 2, 4)

        Inspiration for distance calculation:
            https://github.com/lucidrains/vector-quantize-pytorch/blob/master/vector_quantize_pytorch/vector_quantize_pytorch.py
        """
        assert x.shape[-2:] == (2, self.n_bins), f'Codebook expects (..., 2, {self.n_bins}) got {x.shape}'
        shape = x.shape
        flatten = rearrange(x, '... c bpp -> (...) (c bpp)', c=2, bpp=self.n_bins)
        embed = self.emb.weight.T
        dist = -(flatten.pow(2).sum(1, keepdim=True) - 2 * flatten @ embed + embed.pow(2).sum(0, keepdim=True))
        embed_ind = dist.max(dim=-1).indices
        embed_ind = embed_ind.view(*shape[:-2])
        return embed_ind

    def decode(self, idx: Tensor):
        v = self.emb(idx)
        return rearrange(v, '... (c b) -> ... c b', c=2)

    def forward(self, projection_windows: Tensor):
        return self.encode(projection_windows)


INIT_ARGS = {
    'bin_frames': [0, 1, 2, 3]
}

ARGS = {
    'projection_windows': torch.randn(2, 2, 4)
}

if __name__ == '__main__':
    codebook = Codebook(**INIT_ARGS)
    output = codebook(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output: {output}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 1, dtype family int, shape (2,) — shape (2,) 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
Combinator Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("9a020aa2a0f3c65a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 2, dtype family float, shape (2, 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
import torch.nn as nn

class Combinator(nn.Module):
    """
    Combines the "ego-centric" representations from identical 'towers'
    processing channel 0 and 1. The towers are identical (shared weights)
    and therefore channel agnostic, e.g. they don't know if they process information
    from the view of speaker A or B.

    Here we have specific layers associated with each channel to join the representations
    into a single coherent space with channel information included.
    """

    def __init__(self, dim: int, activation: str='GELU'):
        super().__init__()
        self.dim = dim
        self.h0_a = nn.Linear(dim, dim, bias=False)
        self.h0_b = nn.Linear(dim, dim, bias=False)
        self.ln = nn.LayerNorm(self.dim)
        self.activation = getattr(nn, activation)()

    def forward(self, x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
        """
        Combines the hidden states from identical 'towers' which have processed
        each channel from an 'ego-centric' view. However, the towers are channel agnostic
        by default (shared weights) so in this step we process the information from channel 0, 1
        separately into a joint representation.

        The final representation will (see GPTStereo -> ProjectionModel) go into a final linear
        layer to produce logits.
        """
        ha = self.activation(self.ln(self.h0_a(x1)))
        hb = self.activation(self.ln(self.h0_b(x2)))
        h = ha + hb
        return h


INIT_ARGS = {
    'dim': 4,
    'activation': 'GELU'
}

ARGS = {
    'x1': torch.randn(2, 4),
    'x2': torch.randn(2, 4)
}

if __name__ == '__main__':
    combinator = Combinator(**INIT_ARGS)
    output = combinator(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output: {output}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 (2, 4) — shape (2, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
LayerNorm Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("8679d4611c936d0e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float, shape (2, 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 torch.nn as nn
from einops.layers.torch import Rearrange

class LayerNorm(nn.Module):
    """
    Extending `nn.LayerNorm` by rearranging input dims to normalize over channel dimension in convnets.

    The original `nn.LayerNorm` + 2 einops Rearrange is faster than custom Norm which calculated values directly on channel...
    """

    def __init__(self, dim: int, rearrange_outputs: bool=True) -> None:
        super().__init__()
        self.ln = nn.LayerNorm(dim)
        self.in_rearrange = Rearrange('b d t -> b t d')
        if rearrange_outputs:
            self.out_rearrange = Rearrange('b t d -> b d t')
        else:
            self.out_rearrange = nn.Identity()

    def __repr__(self):
        return str(self.ln)

    def forward(self, x):
        return self.out_rearrange(self.ln(self.in_rearrange(x)))


INIT_ARGS = {
    'dim': 4,
    'rearrange_outputs': True,
}

ARGS = {
    'x': torch.randn(2, 4, 3),
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
einops, 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, 4, 3) — shape (2, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
MultiHeadAttention Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("77041e75fb11ebf7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned tuple. 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
import torch.nn as nn
from einops.layers.torch import Rearrange

INIT_ARGS = {
    'dim': 4,
    'num_heads': 2,
    'dropout': 0.1,
    'bias': False,
}

ARGS = {
    'Q': torch.randn(2, 3, 4),
    'K': torch.randn(2, 3, 4),
    'V': torch.randn(2, 3, 4),
    'mask': None,
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
MultiHeadAttentionAlibi Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("2bdf088dac0aa116")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned tuple. 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
import numpy as np
from scipy import stats

INIT_ARGS = {
    'dim': 4,
    'num_heads': 2,
    'dropout': 0.1,
    'bias': False,
    'context_limit': 8
}

# Create input tensors for Q, K, V with batch=2, seq_len=3, dim=4
Q = torch.randn(2, 3, 4)
K = torch.randn(2, 3, 4)
V = torch.randn(2, 3, 4)

ARGS = {
    'Q': Q,
    'K': K,
    'V': V,
    'mask': None
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
einops, numpy, scipy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
ProjectionWindow Ran
ran on a synthesized input
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("9aee7648ff92125c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 4, 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 torch
import numpy as np

# INIT_ARGS for ProjectionWindow.__init__
INIT_ARGS = {
    'bin_times': [0.2, 0.4, 0.6, 0.8],
    'frame_hz': 50,
    'threshold_ratio': 0.5
}

# ARGS for ProjectionWindow.__call__
# The __call__ method takes va: Tensor with shape (B, N, C)
# Based on the code:
# - projection extracts va[..., 1:, :] and unfolds with size=sum(bin_frames)
# - bin_frames = [10, 20, 30, 40] (from bin_times * frame_hz)
# - horizon = sum(bin_frames) = 100
# - unfold needs at least horizon frames after slicing off first frame
# So N should be at least horizon + 1 = 101

B = 2  # batch size
N = 105  # number of frames (needs to be > horizon + 1)
C = 3  # number of channels

va = torch.rand(B, N, C)

ARGS = {
    'va': va
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
bin_times_to_frames Ran
ran, violated the contract we drafted
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("19873c62eb915c36")
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) and returned list, ndim 1, dtype family int; 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 torch
from typing import List

torch.manual_seed(0)

# bin_times_to_frames takes:
# - bin_times: List[float] - a list of floating point time values
# - frame_hz: int - frames per second (an integer)

# The function multiplies bin_times by frame_hz and converts to long integers
# So we need valid time values and a valid frame rate

bin_times_value = [0.0, 0.5, 1.0, 1.5, 2.0]
frame_hz_value = 30

ARGS = {
    "bin_times": bin_times_value,
    "frame_hz": frame_hz_value,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — 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 int — 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
ensure_model Ran
ran; our contract draft was wrong, not the code
pipecat-ai/smart-turn/record_and_predict.py
code served (permissive licence) · get_code("1176d5658e5240d8")
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 str. 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 os
import tempfile
from unittest.mock import patch, MagicMock

# Create a temporary directory for testing
temp_dir = tempfile.gettempdir()
test_model_path = os.path.join(temp_dir, 'test_silero_vad.onnx')

# Mock the urlretrieve to avoid actual downloads
def mock_urlretrieve(url, path):
    # Create a dummy file instead of downloading
    with open(path, 'w') as f:
        f.write('dummy model')

# Create the test file to simulate that model already exists
# This way we test the case where the model is already downloaded
with open(test_model_path, 'w') as f:
    f.write('dummy model content')

ARGS = {
    'path': test_model_path,
    'url': 'https://github.com/snakers4/silero-vad/raw/master/src/silero_vad/data/silero_vad.onnx'
}
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 str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
ffn_block Ran
ran; our contract draft was wrong, not the code
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("1a33f1ad866748f1")
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) and returned Sequential. 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
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    'din': 8,
    'dff': 16,
    'activation': 'GELU',
    'dropout': 0.1,
    'bias': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Sequential — a Sequential output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Sequential output the recorded ndim 1 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
get_cnn_layer Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("7d1a088bfea20378")
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) and returned Sequential. 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
import torch.nn as nn
from typing import List
from einops.layers.torch import Rearrange

torch.manual_seed(0)

class LayerNorm(nn.Module):
    """
    Extending `nn.LayerNorm` by rearranging input dims to normalize over channel dimension in convnets.

    The original `nn.LayerNorm` + 2 einops Rearrange is faster than custom Norm which calculated values directly on channel...
    """

    def __init__(self, dim: int, rearrange_outputs: bool=True) -> None:
        super().__init__()
        self.ln = nn.LayerNorm(dim)
        self.in_rearrange = Rearrange('b d t -> b t d')
        if rearrange_outputs:
            self.out_rearrange = Rearrange('b t d -> b d t')
        else:
            self.out_rearrange = nn.Identity()

    def __repr__(self):
        return str(self.ln)

    def forward(self, x):
        return self.out_rearrange(self.ln(self.in_rearrange(x)))

def get_cnn_layer(dim: int, kernel: List[int]=[5], stride: List[int]=[2], dilation: List[int]=[1], activation: str='GELU'):
    layers = []
    layers.append(Rearrange('b t d -> b d t'))
    for k, s, d in zip(kernel, stride, dilation):
        layers.append(nn.Conv1d(dim, dim, kernel_size=k, stride=s, dilation=d))
        layers.append(LayerNorm(dim))
        layers.append(getattr(torch.nn, activation)())
    layers.append(Rearrange('b d t -> b t d'))
    return nn.Sequential(*layers)

ARGS = {
    'dim': 4,
    'kernel': [5],
    'stride': [2],
    'dilation': [1],
    'activation': 'GELU'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Sequential — a Sequential output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Sequential output the recorded ndim 1 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
infer Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
ten-framework/ten-turn-detection/inference.py
pointer only (licence: NOASSERTION) · get_code("b3a109b2477a61cc")
How it ran

The body of this sample is withheld: GitHub could not classify the licence of this paper's repository, ten-framework/ten-turn-detection (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

process_predictions Ran
ran, violated the contract we drafted
pipecat-ai/smart-turn/benchmark.py
code served (permissive licence) · get_code("ddab306151b570e3")
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 tuple, ndim 2, 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 numpy as np

# Create a sample logits array
# The function expects np.ndarray
# It uses squeeze(), checks ndim, reshapes if needed, and does comparison with 0.5
logits = np.array([[0.8, 0.2, 0.6, 0.1]])

ARGS = {
    "logits": logits
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 2, dtype family float — a tuple 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
truncate_audio_to_last_n_seconds Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
pipecat-ai/smart-turn/audio_utils.py
code served (permissive licence) · get_code("a8735ade2ccc2274")
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 1, dtype family float, shape (128000,) 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 numpy as np

# Create a small audio array (shorter than 8 seconds at 16kHz)
short_audio = np.random.randn(80000).astype(np.float32)

# Create a long audio array (longer than 8 seconds at 16kHz)
long_audio = np.random.randn(200000).astype(np.float32)

# Test case 1: audio shorter than target duration
audio_array = short_audio

ARGS = {
    "audio_array": audio_array,
    "n_seconds": 8,
    "sample_rate": 16000
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (128000,) — shape (128000,) 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
CPCModel Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("8a810b6dc7287090")
How it ran

no invocation recorded.

EncoderCPC Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("71278103ea6a96ff")
How it ran

no invocation recorded.

GPT Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("3ebb8347295da573")
How it ran

no invocation recorded.

GPTStereo Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("007982a6b6ed944f")
How it ran

no invocation recorded.

ObjectiveVAP Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("ad1ddcb5fae9a2f4")
How it ran

no invocation recorded.

TransformerLayer Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("9ac8298798c97a60")
How it ran

no invocation recorded.

TransformerStereoLayer Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("2b80871d06e1a2d1")
How it ran

no invocation recorded.

Vap Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("e8215714c95b53c2")
How it ran

no invocation recorded.

VapConfig Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("648625ca61f98cce")
How it ran

no invocation recorded.

VapGPT Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("6637b14d54af2d68")
How it ran

no invocation recorded.

VapGPT_bc Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("f74e8ba777b8c7bb")
How it ran

no invocation recorded.

VapGPT_nod Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("36c5342a141012ea")
How it ran

no invocation recorded.

generate_markdown_output_path Not yet run
unverified (harvested, no recorded run)
pipecat-ai/smart-turn/benchmark.py
code served (permissive licence) · get_code("eb31a867e17d2e31")
How it ran

no invocation recorded.

load_CPC Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("ddc8edcb8f1bce43")
How it ran

no invocation recorded.

load_vap_model Not yet run
unverified (harvested, no recorded run)
inokoj/VAP-Realtime/vap_realtime/model.py
code served (permissive licence) · get_code("c26c6001845f37f6")
How it ran

no invocation recorded.

Also present in other papers' code

Of 32 samples lifted from this paper's linked repositories, 32 carry a function identity and 0 do not. Of the 32 distinct functions with an identity, 3 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 1 later) and 29 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.

FunctionAlso present inEarliest-datedOthers named
CPCModel
brick of this paper's assembly: helper
1 other paper
1 dated earlier
1807.03748 (2018-07)
Representation Learning with Contrastive Predictive Coding
bin_times_to_frames
brick of this paper's assembly: helper
1 other paper
1 later
2607.07294 (2026-07)
Multimodal Voice Activity Projection for Turn-Taking in Social Robots with Voice-Activity-Related Pretrained Encoders
ffn_block
brick of this paper's assembly: helper
1 other paper
1 dated earlier
2209.05161 (2022-09)
How Much Does Prosody Help Turn-taking? Investigations using Voice Activity Projection Models

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

Despite recent advances, efficient and robust turn-taking detection remains a significant challenge in industrial-grade Voice AI agent deployments. Many existing systems rely solely on acoustic or semantic cues, leading to suboptimal accuracy and stability, while recent attempts to endow large language models with full-duplex capabilities require costly full-duplex data and incur substantial training and deployment overheads, limiting real-time performance. In this paper, we propose JAL-Turn, a lightweight and efficient speech-only turn-taking framework that adopts a joint acoustic-linguistic modeling paradigm, in which a cross-attention module adaptively integrates pre-trained acoustic representations with linguistic features to support low-latency prediction of hold vs. shift states. By sharing a frozen ASR encoder, JAL-Turn enables turn-taking prediction to run fully in parallel with speech recognition, introducing no additional end-to-end latency or computational overhead. In addition, we introduce a scalable data construction pipeline that automatically derives reliable turn-taking labels from large-scale real-world dialogue corpora. Extensive experiments on public multilingual benchmarks and an in-house Japanese customer-service dataset show that JAL-Turn consistently outperforms strong state-of-theart baselines in detection accuracy while maintaining superior real-time performance.

For agents

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

get_harvested_code_for_paper("2603.26515")
get_code_for_paper("2603.26515")
have("2603.26515")

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

curl https://syntology.ai/api/ran/2603.26515.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.26515.svg)](https://syntology.ai/paper/2603.26515)

Connect an agent — have() is free.