--- title: "Speech and Audio Datasets | Rights-Cleared, Under NDA | YPAI" url: https://ypai.ai/audio/datasets/ description: "Rights-cleared speech and audio YPAI owns or represents, qualified against your device, setting and speakers. A dataset card per item, processed in the EEA." source: "src/copy/routes (route /audio/datasets/)" --- # Speech that already exists. Read by the room it was recorded in. > Rights-cleared speech and audio YPAI owns or represents, qualified against your device, setting and speakers. A dataset card per item, processed in the EEA. - [Speech data](https://ypai.ai/speech-data/) - [Dataset licensing](https://ypai.ai/dataset-licensing/) Rights-cleared speech and audio YPAI owns or represents, qualified against your condition before it is offered. What is missing is collected as new recordings. You describe the speech your system has to hold: the device, the setting, who speaks and how. YPAI reads its own and partner inventory against that condition, shows what can be licensed quickly under NDA, and records only what is missing. Every item arrives as a card your engineers and your counsel can read field by field. - a small booth in an ordinary office - a phone on the ledge, idle - one speaker, recruited by dialect region - consent on record, use set by the schedule ## Four axes describe every item. Choose the ones your model runs under. The same four axes the speech family collects against. Select a value on each and the card below fills for that condition. An axis you leave open stays open on the card, and YPAI collects new recordings for it. - The microphone in the user's hand, not the one in the studio. An item is tagged with the tier it was captured on. - The room, the vehicle, the street. Noise, reverberation and competing speech are recorded on the item, and read on the card. - Who spoke, recruited by dialect region, age and background, so the long tail is represented rather than averaged away. - The speech a benchmark removes and production does not. The frontier asks for overlap and for what happens between the words. Where a required range is not in inventory, YPAI says so at qualification and scopes the collection that fills it. ## An existing item is reviewed against your use before it is offered. Six checks, run by a person who has read your brief, on every item YPAI owns or represents. An item that fails one is set aside. An item that passes is offered with its card. select a value on each axis and the request line fills in ### Intended use - Training, evaluation or a deployed product. The permission differs for each, and the item's licence basis has to carry the one you need. ### Speech type - Read, spontaneous, overlapping, code-switched, expressive. The item's elicitation is read against the speech your system meets in production. ### Acoustic conditions - Setting, device tier and channel, as captured. A studio item is set aside for a cabin deployment and says so. ### Format - Container, channels, transcript schema and metadata levels, checked against what your receiving system reads. ### Provenance - Where the recordings come from, under which agreement, and whether the chain from contributor to licensor is unbroken. ### Permitted rights - What the item may be used for, for how long, where, and whether a derivative model may be deployed. Read on the card, granted by the licence. ## What one dataset card says. Every field a name, every value a word. An example card, in the order buyers read it. Under NDA the same card carries the item's own values; chain of title and provenance gaps are shared under NDA only. example card · no item named - one language, one written norm, the dialect region tagged - the setting as captured, noise and reverberation classed - the speech type as prompted or as it happened - verbatim with disfluencies, speaker turns, word times - the tier it was captured on, one channel per speaker - dialect checked by a linguist from the region - commercial training · evaluation only, held out of training · research - voice cloning, resale of raw audio, white-label supply - per speaker, per purpose; withdrawal quarantines the recordings and is reported - as the schedule sets them, named before signature - prohibited unless the schedule says otherwise A few seconds from your microphone are measured in this tab and discarded. The measurement fills the acoustic fields of the card in words and adds them to the request line. measured in this tab, then discarded the microphone stayed closed · the example stands as written this browser has no microphone path · the example stands as written - the dataset card, in this order - a data card in the order of the EU AI Act's data-governance article - a Croissant description carrying provenance and permitted use - the transcript schema and its guidelines - the QA trail, per dialect group What the card is written against - [The Croissant specification, provenance and permitted use](https://docs.mlcommons.org/croissant/docs/croissant-spec.html) - [Datasheets for Datasets](https://arxiv.org/abs/1803.09010) - [The rights language, ODRL](https://www.w3.org/TR/odrl-model/) - [The consent conditions in the GDPR](https://gdpr-info.eu/art-7-gdpr/) - [The AI Act's data-governance article](https://artificialintelligenceact.eu/article/10/) - [The AI Act's training-content summary duty](https://artificialintelligenceact.eu/article/53/) ## Under NDA you read the inventory. The licence grants the use. The NDA is the first gate and only that. It opens the list. Every use is granted by the licence that follows it. - partner and sourcing information where the agreement allows it Dataset Licence Agreement, or MSA with a SOW and a Data Licence Schedule. - an evaluation licence, held out of training - an exclusive structure, with geography, use or term set Samples are for evaluation. YPAI can ask for samples and evaluation material to be deleted. A named project lead opens the inventory with you. ## Where the inventory has the gap, YPAI records it. An axis with no matching item is collected as new recordings. Recruitment by dialect region, setting and device, on the platform the inventory was built on, so the new recordings arrive with the same card as the licensed ones. Whether an item carries it is shared under NDA. What it lacks is collected new. The industries that ask for this speech - [Automotive, in-cabin speech](https://ypai.ai/solutions/automotive/) - [Healthcare, clinical speech](https://ypai.ai/solutions/healthcare/) - [Finance and contact centres](https://ypai.ai/financial-services-ai-solutions/) - [Industrial and energy](https://ypai.ai/solutions/industrial-energy/) - [Media and platforms](https://ypai.ai/solutions/media-and-platforms/) - [AI companies and voice agents](https://ypai.ai/solutions/ai-companies/) ## Describe the speech you need. Start under NDA. Two sentences are enough. A named project lead reads it, matches it against own and partner inventory, and comes back with what can be licensed and what has to be recorded. A named project lead reads it and answers inside one EU business day. ### Name The speech you need, and what it is for - Existing inventory first, collect only the gap I agree to the privacy policy and to YPAI contacting me about this request. ## Questions ASR and TTS teams ask ### Why is there no list of datasets on this page? The inventory, the dataset cards, the prices and the provenance descriptions are shared under NDA. The page shows the axes an item is described on and one example card, so you can tell before the NDA whether the shape fits. ### What does "YPAI owns or represents" mean? Some items YPAI recorded and controls with sufficient rights. Others YPAI represents for partner suppliers, after sampling, technical review and a rights check. Both arrive with the same card and the same licence structure. ### Can the same item be licensed for training and for evaluation? Only if the schedule says so. Evaluation, model training and commercial deployment are separate permissions, each with territory, term, withdrawal handling, deletion and retention written in. An evaluation set held out of training stays held out. ### What happens when a speaker withdraws consent after delivery? The recordings are quarantined in YPAI systems, the withdrawal is reported to you with a ledger entry, and erasure follows within the period the schedule sets. ### What if no item covers our condition? YPAI says so at qualification and scopes the collection that fills the gap: recruitment by dialect region, setting and device, on the same platform, delivered with the same card. ### Where is the audio processed and stored? Primary operations are in Norway on EEA infrastructure. Items can be structured for EEA-only sourcing and processing; other delivery is arranged where a deployment requires it. Described by the conditions it was recorded in. Describe the condition. The card that matches comes back under NDA; the field that does not is collected as new recordings.