OTS speech datasets, around 35,000 hours ready to license
Our OTS (off-the-shelf) speech datasets are around 35,000 hours of speech, prompts and conversations already recorded by people in our 30,000+ network, roughly 500 hours per language. Every speaker consented to commercial AI training, licence terms are agreed with each customer, and sample clips come back in reply to your request.
Need a language or setting that isn't in stock? We collect speech to your brief too.
What is an OTS speech dataset?
An OTS, or off-the-shelf, speech dataset is speech that's already recorded and ready to license, instead of collected after you place an order. You pick the languages and hours you need from what's in stock, agree the licence, and the data is delivered.
The point is time. A new collection has to find speakers, record them and check every file before anything ships. Speech that's already recorded skips that wait, which is why teams use it to train a first speech recognition model, widen a model to new languages, or build a test set fast.
Our pool holds around 35,000 hours of speech, prompts and conversations, roughly 500 hours per language. People from our 30,000+ network recorded it, and every one of them consented to their voice being used for commercial AI training.
Public corpora are free but often limit commercial use, and they cover what their creators chose to record. Before you build on one, read its licence. Our speech data specs guide covers what to check in any speech dataset.
The speech pool at a glance
- Volume
- Around 35,000 hours, already recorded.
- Per language
- Roughly 500 hours per language in stock.
- Content
- Speech, prompts and conversations.
- Speakers
- People from our 30,000+ network.
- Consent
- Every speaker consented to commercial AI training.
- Licence
- Terms agreed with each customer, in writing.
- Price
- Quoted per order, by languages, hours and licence.
- Samples
- Sample clips sent in reply to your request.
- Security
- ISO 27001 certified. SOC 2 report available under NDA. GDPR DPAs and HIPAA BAAs on request.
- Gaps
- Anything not in stock, collected to your brief.
Speech datasets by language
Each language page lists what's in stock for that language, what to check in a sample and how to license it. Don't see yours? Ask what's in stock.
What teams train on off-the-shelf speech data
Common uses for speech that's already recorded. They show the kind of work, and they aren't client names.
Speech recognition (ASR)
A base of real speech to train or fine-tune a speech-to-text model in a new language, before spending on a custom collection.
Voice bots and IVR
Prompts and conversations that teach a voice bot how people in each language actually ask for things.
Multilingual speech models
Hours in several languages at once, to widen a speech or voice model past the languages it already handles.
Test and evaluation sets
Held-back recordings your model never trained on, to measure word error rate and compare models per language.
Language identification
Speech from many languages side by side, for models that work out which language someone is speaking.
A head start on a custom project
License what's in stock now, then commission only the accents, devices or wording that are missing.
Off-the-shelf speech data or a custom collection?
Both come from the same network, so you don't have to pick a vendor for each. Here's how they compare.
| Topic | Off-the-shelf, from our pool | Custom collection |
|---|---|---|
| When it's ready | Already recorded, so delivery follows the licence | Recorded after the brief is agreed |
| Fit to your product | A broad base in each language | Your accents, devices, noise and wording |
| Volume | Roughly 500 hours per language in stock | As many hours as the brief needs |
| Checking it first | Sample clips sent in reply to your request | A small free sample recorded to your brief |
| Licence | Agreed with you before delivery | Agreed with you before delivery |
| Best for | A first model, new languages, test sets | Production accuracy in a specific setting |
What to ask any speech data vendor, and our answers
Buying speech data goes wrong in the same few places. Ask every vendor these questions. Here's how we answer them.
| Ask this | Why it matters | Our answer |
|---|---|---|
| Did every speaker consent to AI training? | Without it, you can't safely ship a model trained on the data. | Yes. Every speaker consented to commercial AI training. |
| What does the licence let us do? | Some licences cover research or evaluation only, not a commercial product. | Terms are agreed with you in writing, to fit what you're building. |
| Can anyone else license the same recordings? | It decides whether competitors can train on the same data. | It's part of the licence we agree with you. Ask when you request terms. |
| Can we hear it first? | Audio quality and accents only show up when you listen. | Yes. We send sample clips in reply to your request. |
| Is the quote in audio hours or speech hours? | Silence can make up a large share of an audio hour. | We say which one the quote uses. |
| How is our data kept safe? | Your brief and any data you share are confidential. | ISO 27001 certified. SOC 2 report under NDA. GDPR DPAs and HIPAA BAAs on request. |
| What happens if a language is missing? | A second vendor means a second contract and a second format. | We collect it to your brief, from the same network. |
From a stock check to licensed speech data
Your named contact runs it from the first question to delivery.
Ask
Tell us the languages, the hours and what the model is for.
Stock check
We reply with which languages are in stock and how many hours each has.
Sample clips
We send sample clips, so you can hear the audio before you commit.
Licence and quote
We agree the licence terms and send a quote for the order.
Delivery
The licensed speech is delivered to you.
Fill the gaps
Anything missing gets collected to your brief by our network.
Off-the-shelf speech data terms
The words that come up when you license speech data. The full glossary covers every data type.
- OTS (off-the-shelf) dataset
- Data already collected and ready to license before a buyer asks for it, instead of produced to one buyer's brief.
- Custom collection
- New data recorded to one buyer's brief, after they ask for it.
- Scripted speech
- People reading prompts written for them, also called read speech. It gives control over exactly which words get recorded.
- Conversational speech
- People talking naturally with each other, with the pauses, overlaps and corrections of real talk.
- Audio hours and speech hours
- Audio hours count the whole recording. Speech hours count only the parts where someone is talking. Check which one a quote uses.
- Sample rate
- How many times per second the audio is measured, such as 8 kHz for phone calls or 16 kHz for most speech recognition.
- Exclusive licence
- A licence where no one else can license the same recordings. A non-exclusive licence lets the seller license them to others too.
- Speaker consent
- A speaker's recorded agreement to how their voice can be used, which for training data should cover commercial AI training.
Where off-the-shelf speech data leads next
Most teams need one of these once the first model is trained. You can also compare all eight services.
Custom speech data
New speech recorded to your brief, in your setting.
Transcription
Speech turned into timed text, checked by people.
Conversational AI data
Dialogues and intents for voice bots and assistants.
Language directory
100 languages, grouped by region.
Data and model evaluation
Held-back speech scored against your targets.
Questions about OTS speech datasets
Straight answers about licensing speech that's already recorded. The speech data FAQ covers custom collection.
What is an off-the-shelf speech dataset?
Speech that has already been recorded and is ready to license, instead of being collected after you place an order. You pick the languages and hours you need from what's in stock, agree the licence, and the data is delivered.
What does OTS mean?
OTS stands for off-the-shelf. An OTS dataset is data that's already collected and ready to license, so you pick what you need from stock instead of waiting for a new collection.
How much off-the-shelf speech data do you have?
Around 35,000 hours of speech, prompts and conversations, roughly 500 hours per language, recorded by people in our 30,000+ network.
Which languages are in stock?
71 languages are in stock, roughly 500 hours each (Hindi roughly 1,000), and each one has its own page in the list on this page. If a language you need isn't there, we can collect it to your brief.
Did the speakers consent to AI training?
Yes. Every speaker in the pool gave consent for their recordings to be used for commercial AI training.
How is the data licensed?
Licence terms are agreed with each customer, in writing, before delivery. That includes how you can use the data and whether anyone else can license the same recordings, so the terms fit what you're building.
How much does it cost?
We quote each order on its own, based on the languages, the hours and the licence terms. Send what you need and your named contact comes back with a quote.
Can we hear samples before we buy?
Yes. Ask for the languages you need and we send sample clips in reply, so you can check the audio before you commit.
What if our language or domain isn't in the pool?
We collect it. People from our 30,000+ network record new speech to your brief, and many teams license what's in stock first and commission only the gaps.
Should we buy off-the-shelf speech data or commission a custom collection?
Off-the-shelf is faster because the speech is already recorded, and it works well for a first model or a broad base. A custom collection fits better when you need specific accents, devices, noise or wording. Many teams use both.
Do you work with US teams?
Yes. We take calls in US Eastern mornings, which is evening in India. We're ISO 27001 certified, our SOC 2 report is available under NDA, and we sign GDPR data processing agreements and HIPAA business associate agreements on request.
Check what's in stock
List the languages you need, roughly how many hours and what the model is for. We'll tell you what's in the pool and send sample clips.
- Your named contact replies within one business day.
- We tell you which languages are in stock and send sample clips.
- We agree the licence terms and send a quote.
This opens your email app with the request filled in. Nothing is sent until you press send. You can also use the contact page.