SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2305.09652 · 2023

The Interpreter Understands Your Meaning: End-to-end Spoken Language Understanding Aided by Speech Translation

arXiv · PDF · Open in the Atlas

Code that ran

10 of 11 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 9 have no contract verdict; 1 failure 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/2305.09652.json · badge: /api/ran/2305.09652.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
idiap/translation-aided-slu canonical (classifier label)
classifier basis: census
10 ran · 1 no recorded run
FunctionStatusWhere it lives
build_asr_model Ran
ran on a synthesized input
idiap/translation-aided-slu/models/build.py
code served (permissive licence) · get_code("a65d1cf4ea441a9a")
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 Identity. 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 unittest.mock import Mock

torch.manual_seed(0)

class HParams:
    def __init__(self):
        self.asr_model_type = 'none'
        self.adaptor_output_dim = 768
        self.pad_token_id = 0
        self.feat_proj_dropout = 0.1
        self.final_dropout = 0.1
        self.hidden_dropout = 0.1
        self.activation_dropout = 0.1
        self.mask_time_prob = 0.075
        self.remove_top_layers = 0
        self.asr_keep_feature_encoder_only = False

ARGS = {
    'hparams': HParams()
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Identity — a Identity output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Identity 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
build_classifier Ran
ran on a synthesized input
idiap/translation-aided-slu/models/build.py
code served (permissive licence) · get_code("25d5a7d5c1a1d739")
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 Classifier. 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 unittest.mock import Mock

torch.manual_seed(0)

# Create a minimal hparams object with all required attributes
hparams = Mock()
hparams.classifier_type = 'none'
hparams.classifier_head_type = 'classifier'
hparams.classifier_hidden_size = 8
hparams.classifier_num_labels = 3
hparams.classifier_num_targets = 2
hparams.classifier_use_projector = True
hparams.classifier_projector_init_std = 0.02
hparams.classifier_pooling = 'mean'
hparams.classifier_dropout = 0.1
hparams.classifier_keep_layer_from = 0
hparams.classifier_keep_layer_to = 2
hparams.classifier_name = 'xlm-roberta-base'
hparams.qa_segment_cls = False

# Define input_dim
input_dim = 4

ARGS = {
    'hparams': hparams,
    'input_dim': input_dim,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Classifier — a Classifier output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Classifier 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
build_text_decoder Ran
ran on a synthesized input
idiap/translation-aided-slu/models/build.py
code served (permissive licence) · get_code("7fc2ab07411c3b47")
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 MockBartDecoder. 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 transformers import BartConfig, T5Config, MBartConfig

torch.manual_seed(0)

class MockHParams:
    def __init__(self):
        self.decoder_model_name = "facebook/bart-base"
        self.decoder_name = "facebook/bart-base"
        self.decoder_type = "bart"
        self.decoder_dropout = 0.1
        self.decoder_stack_encoder = True
        self.decoder_remove_decoder = False

class MockBartModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.encoder = nn.Linear(4, 3)
        self.decoder = nn.Linear(3, 4)
        self.shared = nn.Embedding(10, 4)
    
    def num_parameters(self):
        return sum(p.numel() for p in self.parameters())

class MockBartDecoder(nn.Module):
    def __init__(self):
        super().__init__()
        self.model = MockBartModel()
        self.lm_head = nn.Linear(4, 10)
        self.config = type('Config', (), {'max_length': 100})()
    
    def num_parameters(self):
        return sum(p.numel() for p in self.parameters())
    
    @classmethod
    def from_pretrained(cls, model_name, dropout=0.1, activation_dropout=0.1):
        return cls()

class MockT5Model(nn.Module):
    def __init__(self):
        super().__init__()
        self.encoder = nn.Linear(4, 3)
        self.decoder = nn.Linear(3, 4)
        self.shared = nn.Embedding(10, 4)
    
    def num_parameters(self):
        return sum(p.numel() for p in self.parameters())

class MockT5Decoder(nn.Module):
    def __init__(self):
        super().__init__()
        self.encoder = nn.Linear(4, 3)
        self.decoder = nn.Linear(3, 4)
        self.shared = nn.Embedding(10, 4)
        self.lm_head = nn.Linear(4, 10)
        self.config = type('Config', (), {'max_length': 100})()
    
    def num_parameters(self):
        return sum(p.numel() for p in self.parameters())
    
    @classmethod
    def from_pretrained(cls, model_name):
        return cls()

import transformers.models.bart.modeling_bart as bart_module
import transformers.models.t5.modeling_t5 as t5_module

original_bart_from_pretrained = bart_module.BartForConditionalGeneration.from_pretrained
original_t5_from_pretrained = t5_module.T5ForConditionalGeneration.from_pretrained

def mock_bart_from_pretrained(cls, model_name, **kwargs):
    return MockBartDecoder()

def mock_t5_from_pretrained(cls, model_name, **kwargs):
    return MockT5Decoder()

bart_module.BartForConditionalGeneration.from_pretrained = classmethod(lambda cls, model_name, **kwargs: MockBartDecoder())
t5_module.T5ForConditionalGeneration.from_pretrained = classmethod(lambda cls, model_name, **kwargs: MockT5Decoder())

from transformers import BartForConditionalGeneration, T5ForConditionalGeneration
BartForConditionalGeneration.from_pretrained = classmethod(lambda cls, model_name, **kwargs: MockBartDecoder())
T5ForConditionalGeneration.from_pretrained = classmethod(lambda cls, model_name, **kwargs: MockT5Decoder())

hparams = MockHParams()

ARGS = {
    "hparams": hparams
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type MockBartDecoder — a MockBartDecoder output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a MockBartDecoder 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
get_audiofile Ran
ran; our contract draft was wrong, not the code
idiap/translation-aided-slu/dataloader.py
code served (permissive licence) · get_code("6bc062e8e9560b15")
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 BytesIO. 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
import zipfile
import io

# Create a temporary directory structure for testing
temp_dir = tempfile.mkdtemp()
sub_dir_name = "audio_data"
sub_dir_path = os.path.join(temp_dir, sub_dir_name)
os.makedirs(sub_dir_path, exist_ok=True)

# Create a test audio file with .flac suffix
test_filename = "sample_001_abc_defg.flac"
test_file_path = os.path.join(sub_dir_path, test_filename)
with open(test_file_path, 'wb') as f:
    f.write(b"fake audio data")

# Also create a zip file alternative for testing the zipfile path
zip_path = os.path.join(temp_dir, sub_dir_name + '.zip')
with zipfile.ZipFile(zip_path, 'w') as zf:
    zf.writestr(test_filename, b"fake audio data from zip")

ARGS = {
    "datadir": temp_dir,
    "sub_dir": sub_dir_name,
    "name": "sample_001_abc_defg"
}
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 BytesIO — a BytesIO 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: none by the harness; any seed is in the setup block
get_audiofile_spec Ran
ran on a synthesized input
idiap/translation-aided-slu/dataloader.py
code served (permissive licence) · get_code("537bc17db309f75e")
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 BufferedReader. 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 io
import os
import zipfile
import tempfile
import shutil

# Create a temporary directory for testing
test_dir = tempfile.mkdtemp()

# Create subdirectories and files for testing
datadir = test_dir
sub_dir = "audio_data"
os.makedirs(os.path.join(datadir, sub_dir), exist_ok=True)

# Create a test audio file
test_filename = "sample_001_file.wav"
test_file_path = os.path.join(datadir, sub_dir, test_filename)
with open(test_file_path, 'wb') as f:
    f.write(b"fake audio data")

# Test case 1: Direct file in subdirectory
ARGS = {
    "datadir": datadir,
    "sub_dir": sub_dir,
    "name": test_filename
}
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 BufferedReader — a BufferedReader 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: none by the harness; any seed is in the setup block
is_weight_decayed Ran
ran on a synthesized input
idiap/translation-aided-slu/models/model.py
code served (permissive licence) · get_code("a0ad3496d9b49bfc")
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 bool, ndim 0, dtype family float, shape () 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

torch.manual_seed(0)

# The function is_weight_decayed takes a parameter 'n' which is a string
# It splits the string by '.' and checks the last and second-to-last parts
# A typical use case would be parameter names from a neural network model

# Create sample parameter names that would be used with this function
param_name_1 = "layer.0.weight"  # Should be weight decayed
param_name_2 = "layer.0.bias"    # Should NOT be weight decayed (ends with bias)
param_name_3 = "layer_norm.weight"  # Should NOT be weight decayed (second-to-last is layer_norm)

# The function takes a single parameter 'n' which is a string
ARGS = {
    "n": "layer.0.weight"
}
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 bool, 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
learning_rate_schedule Ran
ran on a synthesized input
idiap/translation-aided-slu/models/model.py
code served (permissive licence) · get_code("2ae3cdbae25787c9")
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 int, ndim 0, dtype family int, shape () 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)

class HyperparameterConfig:
    def __init__(self):
        self.reset_period = 1000
        self.reset_times = 3
        self.warmup_steps = 100
        self.plateau_steps = 200
        self.decay_steps = 300
        self.decay_type = 'linear'
        self.max_lr = 0.001
        self.min_lr = 0.0001
        self.final_lr = 0.00001

hp = HyperparameterConfig()
global_step = 250

ARGS = {
    'global_step': global_step,
    'hp': hp
}
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 int, ndim 0, dtype family int, 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
normalize Ran
ran on a synthesized input
idiap/translation-aided-slu/corpora/gigawords.py
code served (permissive licence) · get_code("068c70006f05aa82")
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 re

def normalize(text):
    text = text.replace('-lrb-', '(').replace('-rrb-', ')').replace('-lsb-', '[').replace('-rsb-', ']').replace('-lcb-', '{').replace('-rcb-', '}').replace('&', '&').replace('{ ndash }', '--').replace('/ {', '').replace('/ }', '').replace('{ rsquo }', "'").replace('{ lsquo }', "'").replace('{ ldquo }', '"').replace('{ rdquo }', '"').replace('{ emdash }', '--').replace('–', '--')
    text = (' ' + text + ' ').replace(" ' s ", "'s ").replace(" ' t ", "'t ").replace(" ' m ", "'m ").replace(" ' ve ", "'ve ").replace(" ' re ", "'re ").replace(" ' d ", "'d ").replace(" ' ll ", "'ll ")
    text = text.replace(" n't ", "n't ").replace(" 's", "'s").replace(" 'd ", "'d ").replace(" 'll ", "'ll ").replace(" 'm ", "'m ").replace(" 've ", "'ve ").replace(" 're ", "'re ")
    text = text.replace('\\', '').replace('_', ' ').replace('``', "''").replace('`', "'").replace('~', '').replace('^', '')
    for ch in ',.!?:;)':
        text = text.replace(' ' + ch, ch)
    text = text.replace('( ', '(')
    text = text.replace('port-au - prince', 'port-au-prince')
    text = (' ' + text + ' ').replace(' skorea', ' south korea').replace(' nkorea', ' north korea').replace(' aa ', ' AA ').replace(' aaa ', ' AAA ').replace('u.s.', 'US').replace('u.s', 'US').replace('u.n.', 'UN').replace('u.n', 'UN').replace('u.k.', 'UK').replace('u.k', 'UK').replace('at&t', 'AT&T')
    for word in ['un', 'us', 'eu', 'au', 'roc']:
        text = re.sub('\\b%s\\b' % word, word.upper(), text)
    text = text.replace('í', 'i')
    text = text.replace('+ postponed indefinitely +', 'postponed indefinitely').replace('+ speculative +', 'speculative').replace('+ the people +', 'the people').replace('+ iran +', 'iran').replace('+ the star +', 'the star')
    text = text.replace('port-AU-prince', 'port-au-prince')
    while '---' in text:
        text = text.replace('---', '--')
    text = re.sub('\\s+', ' ', text)
    return text.strip().strip('*').strip()

ARGS = {
    'text': "Hello -lrb- world -rrb- ! This is a test with u.s. and & symbols. The text has   multiple    spaces and should be normalized properly."
}
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
normalize Ran
ran on a synthesized input
idiap/translation-aided-slu/corpora/mustc.py
code served (permissive licence) · get_code("6d0ce37650057628")
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 re
import unicodedata

# The function takes a single parameter 'text' which should be a string
# Test with various text inputs to demonstrate the function works

ARGS = {
    'text': "Hello Ō world with ∞ & ′ characters (Swahili) [test]"
}
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
to_long_nonneg Ran
ran on a synthesized input
idiap/translation-aided-slu/models/model.py
code served (permissive licence) · get_code("8208905c59cd78a1")
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 int, 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 numpy as np

torch.manual_seed(0)

# Test case 1: torch.Tensor input
tensor_input = torch.tensor([[-1.5, 0.3, 2.7, -0.1], [1.2, -3.8, 0.0, 4.9]])

# Test case 2: numpy.ndarray input
numpy_input = np.array([[-1.5, 0.3, 2.7, -0.1], [1.2, -3.8, 0.0, 4.9]])

# Test case 3: scalar input (int or float)
scalar_input = 3.7

ARGS = {
    "r": tensor_input
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
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 2, dtype family int, 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: none by the harness; any seed is in the setup block
normalize Not yet run
unverified (harvested, no recorded run)
idiap/translation-aided-slu/corpora/covost2.py
code served (permissive licence) · get_code("2d04e37fb39df9f3")
How it ran

no invocation recorded.

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 distinct functions with an identity, 1 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 1 later) and 10 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
get_audiofile 1 other paper
1 later
2410.06846 (2024-10)
Joint Fine-tuning and Conversion of Pretrained Speech and Language Models towards Linear Complexity

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

End-to-end spoken language understanding (SLU) remains elusive even with current large pretrained language models on text and speech, especially in multilingual cases. Machine translation has been established as a powerful pretraining objective on text as it enables the model to capture high-level semantics of the input utterance and associations between different languages, which is desired for speech models that work on lower-level acoustic frames. Motivated particularly by the task of cross-lingual SLU, we demonstrate that the task of speech translation (ST) is a good means of pretraining speech models for end-to-end SLU on both intra- and cross-lingual scenarios. By introducing ST, our models reach higher performance over baselines on monolingual and multilingual intent classification as well as spoken question answering using SLURP, MINDS-14, and NMSQA benchmarks. To verify the effectiveness of our methods, we also create new benchmark datasets from both synthetic and real sources, for speech summarization and low-resource/zero-shot transfer from English to French or Spanish. We further show the value of preserving knowledge for the ST pretraining task for better downstream performance, possibly using Bayesian transfer regularizers.

For agents

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

get_harvested_code_for_paper("2305.09652")
get_code_for_paper("2305.09652")
have("2305.09652")

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

curl https://syntology.ai/api/ran/2305.09652.json

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

[![Syntology run record](https://syntology.ai/api/ran/2305.09652.svg)](https://syntology.ai/paper/2305.09652)

Connect an agent — have() is free.