Speech Data for AI

Speech Data Collection
for AI Training

Collect voice recordings from people in the countries you choose. Build one task with audio, a typed transcript and speaker details for ASR, TTS and voice AI.

Workers in 200+ countriesAudio and transcripts in one taskReview every recording

200+

Countries

13,487

Tasks completed

195,000+

Registered workers

~4 Hours

Avg time to results

How It Works

How Speech Data Collection Works

Request access to AI tasks once. After approval, you set up and launch speech tasks yourself, without long negotiations.

Person writing task notes in a notebook next to a laptop
01 路 Create

Describe what people should record

Add a script to read, a topic to talk about or a list of voice commands. For transcripts, use Custom Data Collection and add a long text field next to each audio field.

  • Scripts, topics or voice commands
  • Transcript field in a custom form
  • Your own instructions and examples
World map with pins marking selected countries and regions
02 路 Target

Choose countries and speakers

Select the countries you need, set the number of recordings and ask for language, accent or age in your instructions or as custom form fields.

  • Any number of countries in one task
  • Speaker details in instructions or form fields
  • You set the number of recordings
Man recording a voice message on his smartphone
03 路 Record

Workers record on their own devices

Your task becomes available to workers in the selected countries. They record in the browser or upload a file, and type the transcript if your form asks for it.

  • Workers in your target countries
  • Browser recorder or file upload
  • Audio and transcript in one custom form
Woman in headphones reviewing recordings on a laptop
04 路 Review

Approve only the recordings that fit

Listen to each submission, check the transcript and reject anything that does not meet your requirements. Keep a clean dataset for training.

  • Listen to every recording
  • Approve or reject each submission
  • Pay only for what you accept
Speech Data Types

Types of Speech Data You Can Collect

Set up a task for the exact type of recordings your model needs. Workers follow your instructions and record on their own phones.

Scripted Speech

Every recording matches your script

Workers read sentences, phrases or numbers from your script. You know exactly what was said in every recording, so the data is ready for training without guesswork.

  • Your own sentences, phrases or numbers
  • Same script read by many different speakers
  • Text is known in advance for every recording
Create a speech task
Woman reading a printed script aloud
Spontaneous Speech

Natural speech the way people really talk

Workers speak freely on a topic you set, like describing their day or giving an opinion. You get real phrasing, pauses and fillers that scripted data cannot give you.

  • Topics and questions you define
  • Natural pauses, fillers and real phrasing
  • Transcript field in a custom form
Create a speech task
Smiling woman speaking freely into her smartphone
Voice Commands

Commands and wake words from thousands of voices

Workers repeat short commands and trigger phrases in their own voice and manner. Collect the variety your voice assistant needs to respond to anyone.

  • Wake words, commands and short phrases
  • Many repetitions from different speakers
  • Different ages, genders and voices
Create a speech task
Smart speaker waiting for a voice command on a table
Dialogues

Real conversations between two people

Workers record a conversation with a friend or family member, following roles or a free topic. Both participants agree to the recording.

  • Role-based or free conversations
  • Real back-and-forth speech
  • Consent from both participants
Create a speech task
Two women having a recorded conversation with a microphone
Accents and Dialects

Cover the accents your users actually have

Collect the same phrases from speakers in different countries and regions. Your model learns to understand people beyond one standard accent.

  • Speakers from the countries you choose
  • Same phrases across different regions
  • Speakers screened by your language requirements
Create a speech task
Man recording his voice on a smartphone outdoors
Noisy Environments

Speech recorded in real-world conditions

Workers record on the street, in a car, in a caf茅 or at home with background sounds. Your model learns to hear speech where people actually use it.

  • Street, car, caf茅 or home settings
  • Real background noise, not added later
  • Location type as a custom form field
Create a speech task
Woman recording a voice message on a city street at night
Task Example

Recordings With Transcripts
in One Task

Build a custom form with an audio field and a transcript field. Workers send both together, so you skip a separate transcription step.

Your setup

What you set up

Task settings

Choose Custom Data Collection, add an audio field and a long text field for the transcript, then write instructions with phrases to read and recording rules.

  • Custom data collection task type
  • Audio field for each clip
  • Long text field for each transcript
  • Instructions with your script

Your view

What you receive

Submission

Each submission arrives as an audio file with the transcript attached, ready to review and approve.

  • Audio and text in one submission
  • Worker country on every submission
  • Approve or reject each pair
  • CSV or ZIP export
Pricing

Estimate Your Speech Data Collection Cost

You set the pay per task. Adjust the numbers to see what your project will cost, with the 15% platform fee included.

?
?
min
?
$
FAQ

Frequently Asked Questions

You set the pay for each recording, starting from $0.10, plus a 15% platform fee. A speech task needs at least 10 workers, so the smallest run costs $1.15. Your first deposit is at least $20.

Once your access to AI tasks is approved and your task is live, workers can start. The average time to results is about 4 hours. Speed depends on the countries, your pay rate and how complex the task is, so a small test run is the quickest way to check.

Yes. Use Custom Data Collection and add an audio field with a long text field for each clip. Workers record or upload the audio and type exactly what they said, so both arrive in one submission.

Accept workers from all countries or pick continents and individual countries; only people from those countries can take your task. Language, accent, age and gender are not automatic filters: ask for them in your instructions or as dropdown fields in a custom form, then reject recordings that do not match.

Yes. Each worker can complete your task only once, so you get as many speakers as submissions. If you need several phrases from one person, set the number of files per submission.

Workers can upload MP3, M4A, AAC, WAV, WebM, OGG or FLAC files up to 50 MB each. The browser recorder saves compressed audio, so if you need WAV or FLAC, ask workers to upload files and state the sample rate and length in your instructions. You can add up to 30 reference files, such as a sample recording, an image or a short video.

You listen to every submission, check the transcript and approve or reject it. You pay only for what you accept, and for larger projects you can launch a separate task where other workers check recordings and transcripts.