Can a company sell its speech and audio recordings to AI companies?
Often yes, when every speaker agreed to be recorded and the recordings belong to the company. Speech models learn from real people talking in real rooms: accents, interruptions, background noise, people correcting themselves. A company that kept years of dictation, voicemail or recorded meetings holds exactly that, and the work is in checking who spoke and what they agreed to.
Audio a company tends to keep
| Dictation and voice notes | A surveyor dictating a site report, an adjuster recording a claim note in the truck, a doctor dictating a letter. The paired typed version makes the audio far more useful. |
| Voicemail archives | Years of messages left for a front desk or a sales line. Short, noisy and full of names and numbers, so they need the most cleaning. |
| Recorded internal meetings | Weekly operations calls, design reviews and training sessions saved from a meeting tool, often with an automatic transcript beside them. |
| Narration for training and product videos | Scripted voice work by staff or hired voice talent. The script exists, so the text and the sound line up word for word. |
| Field and machine audio | Sound from equipment, vehicles or a shop floor, recorded for maintenance or quality checks. Useful for sound recognition, and usually free of speech. |
What makes one hour of audio worth more than another
Labs care about how much speech is in a file, how clear it is, and whether an accurate transcript sits beside it. An hour of meeting audio with eight people talking over each other teaches different things from an hour of one clear narrator, and each has its use.
Labels raise the value. A note of the language, the accent or region, the setting (car, office, phone line) and the speaker count turns a folder of files into a set someone can search. A company rarely has these labels written down, but it often knows them: the dictation came from field staff in three states, the voicemails came into one office.
Count the audio before anyone asks. Total hours, the number of files, the years covered and the share that has a transcript are the four numbers that come up first.
Every voice belongs to someone
A voice can identify a person, and some laws treat a voiceprint as biometric data with its own consent rules. Before any audio is considered, counsel checks two things for each kind of recording: whether everyone on it agreed to be recorded under the law where it was made, and whether that agreement stretches to a license to an outside party.
Staff recordings are the easiest place to start, because an employer can ask for written consent that names the use. Customer voices on voicemail or calls are harder. Some US states require every party to agree to a recording, and a greeting that said "this call may be recorded for quality" was written for quality, not for training.
Hired voice talent is a contract question. The agreement with the narrator says what the recordings may be used for, and many older ones say nothing about AI.
Cleaning audio before a sample is sent
Names, phone numbers, addresses and account numbers spoken aloud are found in the transcript and then muted or replaced with a tone in the audio at the same time stamps. Speakers who did not consent are cut out, and a file where that ruins the recording is dropped.
A small sample, cleaned and approved by the owner, is what goes out first. The full archive stays on company systems until a license is signed.
Audio that stays out
| Recordings made without the required consent | Where the law needed every party to agree and they did not, the file is left out. |
| Customers who never agreed to an outside use | Their voices are removed, or the recording is dropped when removal leaves too little. |
| Privileged or confidential calls | Calls with lawyers, board discussions and anything under a confidentiality agreement are screened out first. |
| Narration whose contract limits reuse | Voice work stays out until the talent agreement, or a new one, covers the use. |
More on this: how personal details are removed, who carries the risk and how you get paid.
An example, invented for illustration
An engineering inspection firm with forty desk staff and a team of field inspectors has eight years of dictated site reports, each with the typed report the office produced from it. The inspectors signed a recording consent form when the firm started using dictation. A sample of two hundred dictations goes out with client names, street addresses and permit numbers muted in the audio and replaced in the text. Voicemails from homeowners stay out because no consent covers them.
Questions owners ask about this
Are transcripts enough, or do labs want the sound itself?
Both have uses. Text from calls helps language models; speech models need the audio, and they need it most when a checked transcript sits beside it, because the pair shows how speech maps to words.
Does background noise make our recordings worthless?
No. Real noise from cars, job sites and busy offices is part of what speech models have to handle, so noisy audio with a good transcript can be worth more than clean studio sound.
Can we license recordings of former employees?
Only if what they signed covers it. A consent form that named outside use still applies after someone leaves; a general monitoring policy may not, and counsel reads the wording.
Didn't find your question? Ask us directly.
Keep reading
Related kinds of records
- Call recordings: Gong, Zoom and phone-system calls; transcripts before audio, and consent before everything.
- Recordings made with consent: Recordings that people agreed to, with spoken details removed and anything unclear left out.
- Records in other languages: Support, sales and team records in languages other than English, plus paired translations, checked by native speakers.
Industries that keep these records
- Engineering and architecture firms: The review comment and the reply behind a design change, with client drawings and seals kept out.
- Music and audio companies: Stems, mix notes and cue sheets, with composition, recording and voice rights each checked before a track is included.