Resources
Speech dataset pricing: raw, transcribed and diarized audio
Compare speech quotes using accepted audio hours, transcript accuracy, speaker labels and separately scoped annotation work.
Published 2026-10-27 · 5 min read
| Item | Define before requesting a quote |
|---|---|
| Raw audio | Format, conditions and accepted duration |
| Transcript | Review level, text conventions and alignment |
| Speaker labels | Overlap, session IDs and cross-session rules |
| Cost comparison | Licence versus annotation and rework |
Raw speech, corrected transcripts and speaker-timed annotations should not share one unexplained price. Ask what each quoted audio hour includes and what happens when an audio file or its labels fail review.
This guide provides a quote specification, not a published market rate. Public research corpora and commercial licensing proposals have different terms; an available research download does not establish what a private recording owner will be paid.
Define the audio product
Record language, dialect coverage, recording conditions, speaker count, channel structure and native format. Separate read speech, interviews, spontaneous conversation and telephone audio rather than mixing their hours into one unexplained total.
OpenSLR describes LibriSpeech as read English speech derived from audiobooks. That is a useful reminder to compare like with like: read speech is not a substitute for every conversational or far-field requirement.
Specify annotation levels separately
For transcripts, agree text conventions, treatment of hesitations and non-speech, timestamps, quality review and correction responsibility. Machine-generated text should be labeled as such until reviewed to the agreed standard.
Diarization identifies who spoke when; it is not automatically a legal identity. Define anonymous speaker IDs, overlapping speech, speaker changes and whether labels follow speakers across sessions. Avoid attaching unnecessary personal identifiers.
Normalize two supplier quotes
Request licence, transcription, speaker labeling, quality control and delivery as separate lines. Define whether the rate applies to elapsed recording time, speech-only segments or the sum of channel durations. Ask how duplicate and rejected material is removed.
Hypothetical arithmetic: $600 of annotation work over 30 accepted audio hours is $20 per accepted hour for that work. It says nothing about the underlying licence price or a collector’s wage. A lower headline total may omit review or rework.
Use a pilot to resolve uncertainty
Evaluate representative languages, acoustic conditions and overlap before scaling. Agree how errors are sampled and adjudicated, what must be corrected and who pays for rework. Do not claim a universal quality threshold; the application and written brief determine it.
An existing supplier can submit a summary to Fiund. A collection team should use the collector path and describe speaker access and recording capability before recruiting people or purchasing equipment. Neither path guarantees paid work.
Frequently asked questions
Does diarization mean a transcript is included?
No. Request the audio, text and speaker timing deliverables explicitly.
Is a public research corpus a commercial price benchmark?
No. Check its own licence and provenance, and compare the actual domain and deliverables rather than treating availability as a market price.