Speech and audio data for AI
Recordings collected to the speaker mix you specify, then transcribed by native speakers with speaker turns, timestamps and noise tags. You get audio files, transcripts and metadata in your format.
- 00:00.0A
Hi, I'm calling about my order. It hasn't arrived yet.
- 00:03.3B
Sorry to hear that. Please read me the order number. [keyboard]
- 00:06.3A
Yeah, it's four eight two, um, four eight two one three.
- 00:09.2B
[overlap] Thanks. It left our warehouse on Tuesday.
Example verbatim transcript. Select a line to jump to it.
- Collection
- Scripted, spontaneous and two-person recordings
- Transcripts
- Verbatim or clean, with speaker turns and timestamps
- Audio specs
- The sample rate, bit depth and channels you set
- Consent
- A signed record for every speaker
What speech data covers
Speech data is recorded audio paired with an accurate transcript and labels. It's used to train and test speech recognition, voice assistants and call analytics. We record new audio to your spec, or transcribe and label audio you already have.
This page is about training data. If you need transcripts for people to read, such as interviews or meeting notes, see transcription.
Recording and labeling work
Mix and match. Most projects combine new recordings with a transcription and labeling pass.
Scripted recordings
Speakers read prompts you supply, such as commands, product names or phonetically balanced sentences, in the language and accent you need.
Conversations and free speech
Talk on set topics, and two-person conversations such as role-played support calls, with each speaker on a separate channel if you need it.
Wake words and commands
Many repetitions of a wake word or short command from many speakers, at different distances from the microphone and in different noise levels.
Transcription for training
Verbatim transcripts that keep fillers, false starts and repetitions, or clean transcripts that follow your rules for numbers, spelling and punctuation.
Speaker and timing labels
Speaker turns, segment and word-level timestamps, overlapping speech, and the points where a speaker switches language mid-sentence.
Noise and event tags
Background noise, music, laughter, coughs and other non-speech events tagged from a label set you approve. Personal details can be marked or bleeped.
You set the speaker mix
Tell us who should be in the data. We recruit to those quotas and report progress with every batch, so gaps show up early.
- Age band
- Gender
- Region and accent
- First or second language
- Phone, laptop or headset
- Quiet room, street or car
- Women, 18 to 34
- 40 / 40
- Women, 35 and over
- 31 / 40
- Men, 18 to 34
- 40 / 40
- Men, 35 and over
- 26 / 40
- Northern English accents
- 38 / 50
- Recorded on a phone
- 60 / 80
Example quota sheet. Your own quotas, speaker attributes and targets go in the project brief.
How a speech data project runs
You hear the pilot recordings and read the transcripts before full collection starts.
Spec
You set the languages, speaker quotas, scripts or topics, audio format and transcription rules.
Pilot
We record and transcribe a small set so you can check the audio quality and transcript style.
Review
You approve the pilot, and we lock the guidelines and recording setup before scaling up.
Your sign-off QACollection
Speakers record under the agreed setup. A second listener checks each transcript against the audio.
Delivery
Audio, transcripts and metadata arrive in batches, with consent records and a batch report.
Audio and file formats
- Audio
- WAV or FLAC, from 8 kHz for telephony up to 48 kHz, 16 or 24-bit, mono or one channel per speaker
- Transcripts
- JSON, CSV, Praat TextGrid, plain text or SRT, with segment or word timestamps
- Metadata
- Speaker ID, age band, gender, region, device and setting for each file, in CSV or JSON
- Tools
- Your annotation platform, or ELAN, Praat or Label Studio set up for your project
Formats and tools to confirm
Consent and privacy
Voice recordings are personal data. These rules apply to every speech project. See how we handle data.
- Signed consent. Every speaker signs a consent form before recording that names how the recordings will be used. You get the consent records with the data.
- Anonymous speaker IDs. Files carry a speaker code, not a name. Personal details said aloud can be marked or bleeped.
- Yours to keep. You own the recordings and transcripts. We don't reuse them or sell them as datasets.
- Deleted at the end. We delete our copies when the project ends, unless you ask us to keep them for follow-up batches.
Related services
Questions about speech data
What is speech data?
Speech data is recorded audio paired with a transcript and labels, such as who is speaking, when each word starts and what background sounds are present. It's used to train and test speech recognition, voice assistants and call analytics models.
Can you collect speech to a set speaker mix?
Yes. You set quotas by age band, gender, region or accent, device and recording setting, and we recruit speakers to fill them. Each batch report shows how far each quota has been filled.
Which audio specs do you deliver?
Whatever your model needs. Common choices are 16 kHz, 16-bit WAV for speech recognition, 8 kHz for telephony, and 44.1 or 48 kHz for higher-quality voice work, in mono or with one channel per speaker.
Do you label speaker turns, accents and background noise?
Yes. We mark speaker turns and overlaps, add timestamps at segment or word level, record each speaker's accent and region in the metadata, and tag noise and other non-speech events from a label set you approve.
How do you get consent from speakers?
Every speaker signs a consent form before recording. It names the purpose of the recordings and how they'll be used, and you receive the consent records with the data.
How is this different from your transcription service?
Transcription is for people to read, so it's formatted for readability. Speech data is for training models, so transcripts follow strict rules for fillers, numbers and timing, and come with speaker metadata and labels.
Start with a pilot set of recordings
Send your spec or book a call. We record and transcribe a pilot so you can check the audio and transcripts before full collection.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time