TCLResearchEurope/intelligent_wakeup

Intelligent Wakeup

A synthetic corpus for device-directed speech detection: multi-speaker conversations in which most speech is not addressed to the voice assistant, with the moments that are clearly marked.

Conventional assistants detect a wake word but cannot tell whether what follows is meant for them. This corpus is built to train and evaluate the module that makes that decision from the whole session, not from an isolated command.

• Project page: https://tclresearcheurope.github.io/intelligent_wakeup/
• Code: https://github.com/TCLResearchEurope/intelligent_wakeup
• Paper: Training Intelligent Voice Assistant Wakeup with Controllable Synthetic Conversations — link TBA

Available versions

Version	Released	Hours	Conversations	What changed
v1.0.0	TBA	TBA	TBA	First public release. As described in paper.

main always points at the newest release. To load a specific one:

from datasets import load_dataset

ds = load_dataset("TCLResearchEurope/intelligent_wakeup", revision="v1.0.0")

To see every version that exists, without downloading anything:

from huggingface_hub import list_repo_refs

refs = list_repo_refs("TCLResearchEurope/intelligent_wakeup", repo_type="dataset")
print([t.name for t in refs.tags])

---

Structure

One row per conversation. An always-on wakeup system is judged on false accepts per hour of continuous audio, which cannot be recovered once sessions are cut into utterances. Per-turn text and timings travel inside the row, so you can explode to utterances yourself.

Field	Type	Description
audio	Audio	The full conversation, mono, 16-bit PCM
id	string	category/variant_type/variant_name
category	string	Scenario domain, e.g. KitchenConversations
variant_type	string	couple (multi-speaker) or single
variant_name	string	Variation id; _long marks an extended session
context	string	One-line scene description
has_va	bool	Whether the assistant is addressed at all
num_turns	int	Turn count
duration_seconds	float	Audio duration
speakers	list[string]	Speaker names in the conversation
turns	list[{speaker, content, time}]	Transcript with onset in seconds
group_key	string	Split group — see the warning below
split	string	train / validation / test

Splits are train / validation / test, assigned by group and stratified by category.

Do not re-split on id. A scenario's short variation, its _long sibling and its no_va_ twin are near-duplicates sharing cast, voices and room tone. They share a group_key, and the provided splits keep them together. Re-splitting per row would place effectively-seen audio in your test set and inflate your results. Group on group_key.

---

Usage

from datasets import load_dataset

ds = load_dataset("TCLResearchEurope/intelligent_wakeup", revision="v1.0.0")
row = ds["train"][0]

row["audio"]["array"]          # waveform
row["has_va"]                  # is the assistant addressed in this session?
for t in row["turns"]:
    print(f'{t["time"]:7.1f}s  {t["speaker"]:10} {t["content"]}')

Decoding audio requires torchcodec with datasets>=4.0:

pip install "datasets>=4.0" torchcodec

To stream instead of downloading everything:

ds = load_dataset("TCLResearchEurope/intelligent_wakeup",
                  revision="v1.0.0", split="train", streaming=True)

Negative-only sessions, useful for measuring false accepts per hour:

neg = ds["test"].filter(lambda r: not r["has_va"])
hours = sum(neg["duration_seconds"]) / 3600

---

Generation

LLM agents write each speaker's dialogue against a phase-structured scenario outline; speech is synthesised with character-specific voices from a pool of 150+ designed voices; room acoustics and generated ambience are applied. Full pipeline and configs are in the repository.

Limitations

• Fully synthetic. Text is LLM-written and speech is TTS. Absolute numbers will not transfer to real recordings; treat it as training and comparative-evaluation material.
• English only.
• Voices are synthetic, so accent, age and pathology coverage reflect the TTS pool rather than a real population.

License

CC BY-NC 4.0 — non-commercial use with attribution. The pipeline code is Apache-2.0.

Citation

@inproceedings{sowanski2026intelligentwakeup,
  title={Training Intelligent Voice Assistant Wakeup with Controllable Synthetic Conversations},
  author={Sowa{\'n}ski, Marcin and Leszczy{\'n}ski, Kacper and Krzywicki, Kacper and Wodnicki, Krzysztof},
  year={2026},
}