Call center data
Role-played customer service calls between native speakers, with agent and customer on separate channels. See call center data.
Custom collection
Custom speech data collection, built to your spec
When our off-the-shelf catalog doesn’t cover your language, domain, or acoustic conditions, our contributor network collects the voice data your model needs, from wake words in far-field noise to hours of in-domain call-center dialogue.
How it works
Nine steps from your brief to your dataset
Andreas walks through each one, from choosing and vetting speakers to transcription and delivery. Jump to any step.
Process
From model requirement to delivered dataset
1 · Spec
We turn your model requirements into a collection spec: language and dialect, speaker demographics, devices, environments, speech type, transcript format, and metadata schema.
2 · Recruit
Vetted native speakers are recruited against your demographic quotas, from age and gender to region and accent, with explicit commercial-use consent captured up front.
3 · Record
Sentence or prompt lists, monologues, or spontaneous conversation, recorded on the devices and in the environments your product will actually encounter: phone, headset, far-field, in-car. For voice cloning and similar projects, you can approve each speaker from a short sample.
4 · QA
Every session passes automated checks (SNR, clipping, duration) plus human review. Sessions that miss spec are re-recorded, not shipped.
5 · Transcribe
Time-aligned, human-verified transcripts in your convention, verbatim or clean-read, with disfluency, overlap, and code-switch tagging as required.
6 · Deliver
Structured delivery with audio, transcripts, per-speaker metadata, and full consent documentation, in the folder layout and format your pipeline expects.
How speakers are vetted
Each speaker passes an ID check and a microphone and hardware test, we measure the background noise where they record, and we confirm their language or dialect. Before recording, they watch a briefing on how the recordings will be used and read the contract for that project, then sign or drop out.
Catalog or custom?
Our off-the-shelf datasets are spontaneous two-person conversations in 60 languages and dialects, licensed as they are at $60 to $95 per hour. When you need scripted speech, chosen voices, a device or a language they don’t cover, we record it to your spec. Read off-the-shelf vs custom speech data.
Common projects
Specialist projects
TTS and voice cloning
Sentence lists and monologues read by native speakers, with voices you can approve from a short sample. See TTS data collection.
Wake words and commands
Your wake phrase, commands and phrases that sound like it, recorded on the devices and in the environments your product will meet. See wake word data collection.
Annotation of your audio
Transcription, timestamps, speaker turns and labels on recordings you already have. See annotation services.
South African languages
isiZulu, isiXhosa and Sesotho recorded by native speakers, plus South African English off the shelf. See South African speech data.
Languages we don't stock
Speakers of low-resource languages and dialects outside the catalog, recruited after a feasibility check. Read the low-resource language guide.
Start a project
Tell us what your model needs
Share your target languages, hours, speaker profiles, and timeline. We’ll come back within one business day with a feasibility assessment and a quote.
Say whether your target is accepted audio hours, utterance counts or recording sessions. Include the quality checks and transcript format that define an accepted delivery. See what drives a custom quote, or book a call.
FAQ
Frequently asked questions
Can you record a custom speech dataset with native speakers in a specific language?
Yes. We recruit vetted native speakers for the language, dialect and region in your brief, against your quotas for age, gender and accent, and record scripted prompts or spontaneous conversation on the devices and in the environments you specify. Feasibility is confirmed for each language before the project starts.
Which languages can you collect?
Every language and dialect in our catalog of 60, and languages outside it, such as isiZulu, isiXhosa and Sesotho. For any other language, we confirm recruitment feasibility before quoting.
How is speaker consent handled?
Every speaker gives explicit commercial-use consent before recording, and the consent documentation is delivered with the dataset for your legal review.
Can you record scripted prompts as well as conversation?
Yes. We record sentence or prompt lists, monologues and spontaneous dialogue between two speakers. Our off-the-shelf catalog is spontaneous conversation only, so scripted speech is always recorded to your spec.
How is custom collection priced?
Per project, based on the language, speaker count and quotas, speech type, devices and environments, transcription depth and timeline. Off-the-shelf datasets have published prices per hour; custom collection is quoted once we have seen your spec.
How quickly do we get a quote?
Within one business day of receiving your brief, we come back with a feasibility assessment and a quote.