Is MP3 good enough to sell?

It can be listed, but expect it to count against you. Lossy compression throws away detail that cannot be recovered, so most speech buyers ask for the original WAV or FLAC masters. If the MP3 is all you have, say so up front.

Do I have to transcribe everything first?

No. Untranscribed audio is still sellable, and some buyers transcribe it themselves. Human-checked transcripts with timestamps make a set more useful, so they tend to widen the list of interested buyers.

What if hosts and guests are on one track?

That is normal for podcasts. It means a buyer has to separate the speakers before it can label them, which is extra work on its side. Separate per-speaker stems, if you kept them, are worth mentioning.

Does background music ruin a recording?

Music under speech makes the audio harder to use for most speech tasks, and the music itself carries its own copyright. Music-free stems or a clean dialogue track are much easier to license.

How do I count hours?

Count speech, not file length. An hour-long episode with a two-minute intro and long musical breaks may hold 50 minutes of usable speech. Buyers will measure it themselves, so a careful estimate saves surprises.

Voice dataset specs: what buyers ask for

Buyers of voice data read a spec sheet before they listen to anything: sample rate, bit depth, channels, speaker count, accents, hours of speech, transcripts and evidence of consent. Clean, lossless, single-speaker tracks with signed releases are the easiest to license; anything missing from that list narrows who will take it.

What goes on a voice dataset spec sheet?

Fill in the right-hand column for your own recordings before you talk to anyone. If you cannot answer a line, that is the first thing to find out.

Field	What buyers look for	Why it moves the value
Sample rate	The rate the masters were recorded at: 44.1 or 48 kHz for studio work, 16 kHz for most speech recognition, 8 kHz for phone audio.	Detail lost at recording time cannot be added back, so a higher native rate keeps more uses open.
Bit depth	16-bit or 24-bit PCM in a lossless file (WAV or FLAC).	Lossy files (MP3, AAC) have already discarded detail; buyers often treat them as a lower grade.
Channels	One speaker per track where possible; stereo mixes of several voices need separating.	Per-speaker tracks save the buyer the work of splitting and labeling voices.
Speakers	How many distinct voices, and the hours each contributes, by anonymous ID.	One voice for many hours suits voice cloning; many voices suit recognition. The mix decides who buys.
Accents and languages	First language and accent or region per speaker, self-reported.	Under-represented accents and languages are scarcer than standard US English.
Hours	Hours of speech, not hours of files, with music and silence taken out.	Volume matters, but only once the rest of the sheet passes.
Transcripts	Whether text exists, whether a person checked it, and whether it has timestamps and speaker labels.	Aligned, checked text makes audio usable for training without a labeling project first.
Environment	Studio, home office, phone line or field; noise, echo and any music under speech.	Consistent, quiet audio is cheaper to use. Mixed conditions should be described as separate sets.
Releases	A signed release or contract from every speaker that covers licensing for AI training.	No release, no deal: most buyers will not take a voice they cannot prove consent for.

Studio or phone quality: which matters more?

Both have buyers, for different jobs. Studio recordings at 44.1 or 48 kHz suit voice synthesis and high-quality speech models, where every detail of the voice counts. Phone audio is narrowband: classic landline audio carries only 8 kHz sampling, so the top of the voice is gone. It is worth less for synthesis but is what speech recognition for call handling and IVR has to cope with.

What hurts every set is inconsistency. A podcast that moved from a USB microphone in 2016 to a treated room in 2021 is really two datasets; describe them separately.

What a clean recording means to a buyer

No clipping, and a low, steady noise floor.
Little room echo, so the voice sounds close to the microphone.
No music or sound effects under speech.
Original masters rather than files re-encoded for streaming.

Our hours calculator turns these answers into a short spec line and shows what buyers publish about rates.

What speaker details should you record?

Buyers want to know who is speaking without knowing their name. A typical speaker sheet holds an anonymous ID, age band, gender as the speaker describes it, first language, accent or region, and the hours each speaker contributes. Collect these only with the speaker's agreement, and never ship names or contact details with the audio.

Balance matters as much as volume. Three hundred hours from one narrator is a different product from three hundred hours spread across two hundred voices. Neither is better; they suit different models, and your spec should say which one you have.

Consent is part of the spec too: speaker releases explains what a buyer will expect to see.

Questions about voice specs

Is MP3 good enough to sell?: It can be listed, but expect it to count against you. Lossy compression throws away detail that cannot be recovered, so most speech buyers ask for the original WAV or FLAC masters. If the MP3 is all you have, say so up front.
Do I have to transcribe everything first?: No. Untranscribed audio is still sellable, and some buyers transcribe it themselves. Human-checked transcripts with timestamps make a set more useful, so they tend to widen the list of interested buyers.
What if hosts and guests are on one track?: That is normal for podcasts. It means a buyer has to separate the speakers before it can label them, which is extra work on its side. Separate per-speaker stems, if you kept them, are worth mentioning.
Does background music ruin a recording?: Music under speech makes the audio harder to use for most speech tasks, and the music itself carries its own copyright. Music-free stems or a clean dialogue track are much easier to license.
How do I count hours?: Count speech, not file length. An hour-long episode with a two-minute intro and long musical breaks may hold 50 minutes of usable speech. Buyers will measure it themselves, so a careful estimate saves surprises.