Aslema-Synth-TN

Aslema-Synth-TN is a fully synthetic Tunisian Derja corpus for spoken language understanding: speech annotated for intent and for slot filling. It was built for NADI 2026 Shared Task 5 to cover the intents and slots that are rare or absent in the real SLURP-TN training split, and it is the augmentation set behind the Aslema system, which ranked 1st in slot filling on the official test set.

No human was recorded for this dataset. An LLM wrote the utterance text, a rule filter and a panel of LLM judges cut it down, and a text-to-speech model spoke it.

At a glance

Dataset Attribute	Information
utterances	22,940
audio	22.12 hours, 16 kHz mono
median utterance	3.36 s (range 0.64 to 14.40 s)
intent labels	21
slot types / spans	26 / 23,841
utterances with no slot	6,765 (29%), by design
splits	one train split

Each text was spoken twice, by two versions of the same TTS model. Both are released, marked in the tts_system column.

tts_system	Utterances	Hours	Key
voxcpm_base	11,871	10.38	stock VoxCPM
voxcpm_lora_ft	11,069	11.74	VoxCPM after LoRA fine-tuning on ~2.8 h of SLURP-TN training speech
total	22,940	22.12

The two systems saw the same text and the same reference-voice pool, so the pair supports a direct comparison of TTS fine-tuning as an augmentation choice. The counts differ because each system's output was filtered independently.

Counts in the figure are rounded, and its two per-system numbers are clip counts before the acoustic filter. The released counts are in the table above.

Development pipeline

The above figure summarizes the pipeline used to construct this dataset, from seed utterance selection to final audio rule-based filtering.

Repository contents

• data/train-*.parquet: the train split, 12 shards (2.4 GB), with embedded audio the Hugging Face viewer can play.
• data/metadata.csv: the same rows as a flat table, no audio bytes. Its file_name column names the WAV each row would occupy under data/audio/.
• build_release.py: the script that produced both from the pipeline's working directory, including the anonymization steps.

Loose WAV files are not stored here. 22,940 small LFS objects would make the repo slow to clone for no extra information, since every byte is already in the parquet. Write them out locally in one pass:

import soundfile as sf, pathlib
from datasets import load_dataset

out = pathlib.Path("data/audio"); out.mkdir(parents=True, exist_ok=True)
for row in load_dataset("<your-username>/Aslema-Synth-TN", split="train"):
    sf.write(out / f"{row['id']}.wav", row["audio"]["array"], 16000)

Quick start

from datasets import load_dataset

ds = load_dataset("<your-username>/Aslema-Synth-TN", split="train")
print(ds[0]["intent"], ds[0]["text"])
print(ds[0]["audio"]["sampling_rate"], ds[0]["audio"]["array"].shape)

To read only the labels, without downloading 2.4 GB of audio:

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="<your-username>/Aslema-Synth-TN",
    repo_type="dataset",
    filename="data/metadata.csv",
)
print(pd.read_csv(path).head())

Schema

column	description
id	unique utterance id: synth_<mode>_<intent>_<batch>_<index>, with a trailing _ft for the LoRA-fine-tuned system
audio	synthesized speech, 16 kHz mono
text	Tunisian Derja transcript
annotated	the same transcript with inline SLURP-TN slot markup
intent	intent label, one of 21
slots	JSON string: a list of {"label": ..., "value": ...}
generation_mode	coverage, paraphrase, or distractor
generator_model	the LLM that wrote the utterance
tts_system	voxcpm_base or voxcpm_lora_ft
duration	seconds

data/metadata.csv carries the same columns plus file_name, and drops audio. One row, as it appears there:

file_name,id,text,annotated,intent,slots,generation_mode,generator_model,tts_system,duration
data/audio/synth_cov_addcontact_0_000.wav,synth_cov_addcontact_0_000,بلاهي حط النيمرو هذا في الريبيرتوار,بلاهي حط النيمرو هذا في <list_name> الريبيرتوار >,addcontact,"[{""label"": ""list_name"", ""value"": ""الريبيرتوار""}]",coverage,gemini-3.1-pro-preview,voxcpm_base,2.4

The markup in annotated follows SLURP-TN: a slot opens with <label> and closes with a bare >. Stripping the markup returns text exactly.

Generation modes

generation_mode	utterances	hours	purpose
coverage	17,897	17.35	fresh utterances for an intent, with newly invented slot values; rare intents got a larger quota
distractor	3,357	3.06	slot-free utterances for slot-bearing intents, to hold slot precision up
paraphrase	1,686	1.71	dialectal rephrasings of a real training utterance, with slot labels and values frozen
total	22,940	22.12

Slot values in coverage rows are new, never copied from the training split, so the set also tests robustness to unseen values.

Label inventory

intent	utterances	intent	utterances
general_greet	2,445	email_query	1,173
email_addcontact	2,079	news_query	1,093
general_joke	1,967	email_sendemail	1,058
alarm_query	1,958	weather_query	1,038
alarm_remove	1,824	quirky	350
takeaway_query	1,695	query	336
takeaway_order	1,538	sendemail	285
email_querycontact	1,400	greet	121
alarm_set	1,174	joke	98
general_quirky	1,174	addcontact	74
		querycontact	60
		total	22,940

The label set is the shared task's own, which is why both general_greet and a bare greet appear. They are distinct labels in SLURP-TN, not a normalization error here.

slot	spans	slot	spans	slot	spans
date	4,774	time	2,695	person	2,151
relation	1,742	personal_info	1,586	timeofday	1,421
food_type	1,408	order_type	1,204	place_name	1,079
joke_type	956	business_name	883	event_name	880
business_type	707	weather_descriptor	547	email_address	376
news_topic	375	media_type	284	house_place	199
alarm_type	129	email_folder	127	meal_type	112
general_frequency	92	list_name	84	transport_type	14
personal	12	app_name	4	total	23,841

Intended use

Recommended:

• augmenting low-resource Tunisian Derja SLU training sets. Mixing all 22,940 rows with the 2,677 real SLURP-TN training utterances gives the 25,617-utterance training set used by the Aslema system
• studying LLM-plus-TTS synthetic augmentation for spoken language understanding
• comparing stock against task-fine-tuned TTS, using the tts_system column
• robustness work on unseen slot values and rare intents

Not recommended:

• as an evaluation benchmark. This is augmentation data and its labels came from models, not humans; evaluate on the real SLURP-TN devtest or the shared task's official test set instead
• speaker identification or biometric profiling
• re-identifying, impersonating, or further cloning the voices underlying the audio
• any deployment where a synthetic annotation error would carry a safety cost

Ethics and limitations

The reference clips used for voice cloning come from the train split of the publicly released Elyadata/SLURP-TN corpus, the same corpus the shared task's real training data comes from. No audio recorded outside that public corpus was used, and the mapping from a synthetic clip back to the recording it was cloned from is deliberately not part of this release. Nothing here identifies the machines, accounts, or file paths used to produce it.

An LLM wrote every utterance from a label name and training-split few-shot examples, and was never shown personal data. Slot values are therefore invented entities: names, contact-list names, and email-like strings that fill the SLURP-TN slot schema. As a precaution, any string matching an email address or a run of seven or more digits was replaced before release with contact@example.com and 0000000 respectively; 363 email placeholders remain in the text. No other free-text scrubbing was applied.

Known limits:

• No human annotator ever validated a label. The filter is three automatic voters, and its error rate on this data is unmeasured.
• Synthesized speech lacks the prosody, disfluency, and channel variation of real spontaneous speech, so a model trained on this alone underperforms one trained on the mix. The paper reports both.
• Voice diversity is bounded by the 152-utterance reference pool.
• The intent distribution is deliberately flatter than the real one and reflects no natural prior over user requests.
• Tunisian Derja code-switches with French; the transcripts keep the Latin-script French tokens as generated.

License

Released under CC BY-NC-SA 4.0.

On lineage: the upstream Elyadata/SLURP-TN corpus, whose train split provided the voice-cloning references, declares no license on its dataset card. These terms cover the synthetic material produced here and grant no rights over that upstream corpus.

Citation

@inproceedings{shafiq2026aslema,
    title     = "Aslema at {NADI} 2026: Data Augmentation for Intent Recognition and Slot Filling",
    author    = "Shafiq, Tajwaar and Bhatti, Hunzalah Hassan and
                 Alam, Firoj and Chowdhury, Shammur Absar",
    booktitle = "Proceedings of the Fourth Arabic Natural Language Processing Conference (ArabicNLP 2026)",
    year      = "2026",
    address   = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics"
}

Please also cite the corpus the references and labels derive from, and the TTS model:

@article{elleuch2026slurptn,
    title   = "{SLURP-TN}: Resource for {T}unisian Dialect Spoken Language Understanding",
    author  = "Elleuch, Haroun and Mdhaffar, Salima and Est{\`e}ve, Yannick and Bougares, Fethi",
    journal = "arXiv preprint arXiv:2603.21940",
    year    = "2026",
    url     = "https://arxiv.org/abs/2603.21940"
}

@article{zhou2026voxcpm2,
    title   = "{V}ox{CPM}2 Technical Report",
    author  = "Zhou, Yixuan and Zeng, Guoyang and Liu, Xin and others",
    journal = "arXiv preprint arXiv:2606.06928",
    year    = "2026",
    url     = "https://arxiv.org/abs/2606.06928"
}