Transcription and Speech Data Jobs: Turn Audio Into Training Data

A speech model learns from pairs of audio and text, and the text has to match the audio exactly, not what the speaker meant to say. So the work on this page is verbatim: every false start, repeated word, "um", half-finished sentence and overlapping speaker, marked the way the project's guideline says. Clean transcription, the kind a journalist wants, is a different product and usually a different listing.

Four task types cover most listings. Verbatim transcription from scratch, with timestamps. Speaker diarization, labeling who is talking and where turns start and end. Correction of machine output, where an automatic speech recognition (ASR) system has produced a draft and you fix it, which is faster per minute and is becoming the common form. And speech AI evaluation, where you listen to what a speech model produced or transcribed and grade it. Listings also cover document transcription (reading handwritten or PDF text into a structured format), which uses the same conventions without the audio.

The language in the title decides whether you are eligible. Almost every listing names a language and often a region (Vietnamese as spoken in Taiwan, Portuguese as spoken in Brazil), because dialect and accent are what the model is being trained to handle. Voice recording, where you are the speaker rather than the transcriber, is listed separately on voice and audio.

An original example, written for this page. The clip is ninety seconds of two people arranging a delivery over the phone. There is a dog barking in the background, the speakers talk over each other twice, one of them reads a reference number as "double oh seven", and the call drops for three seconds.

Your guideline says: verbatim, tag non-speech events in square brackets, mark overlap with a defined symbol, write numbers as spoken, and never guess an inaudible word. The output is text with speaker labels and timestamps.

The thing that fails people is guessing. A plausible word written over an inaudible one looks fine to a human reader and is poison to a model, and quality review compares your text against the audio word by word. The second thing is inconsistency: the same event tagged two ways in one file.

You need native or near-native command of the listed language and its regional variety, typing speed that makes verbatim work pay, and the patience to apply a long guideline the same way for hours. Transcriptionists, subtitlers, linguists and language teachers do well, as do bilingual people who have done any kind of annotation work. Clinical and legal transcription experience helps with the specialist listings.

Speech AI evaluation and the engineering-flavored listings (ASR, speech data linguistics) want more: a linguistics or speech technology background, sometimes a degree, and comfort with the tooling. Those are a minority of the roles here and pay at the top of the page's range.

It does not suit people who want the highest hourly rate on the board. This page's pay block is usually the lowest of the four kind-of-work hubs, because verbatim transcription is priced near other language work rather than near expert evaluation. Where it pays well is per-minute rates in a less common language, where a fast, accurate transcriptionist with the right dialect has little competition.

Expect a short audio sample to transcribe under the project's rules, scored on word accuracy and on whether you used the tags correctly, and a language or dialect check that may be a conversation. Listings on this page rarely describe an interview; where they do, it is a review of your delivered sample. Equipment questions are simple: headphones, a quiet room and a reliable connection are the usual requirements, and foot pedals are not expected.

Before the test, read the guideline twice and transcribe a minute of any recording the strict way, tagging everything. The Academy module Your First Annotation Task is about text rating rather than audio, but its lesson on applying a guideline consistently is the one transcription tests measure, and How AI Training Jobs Work explains how quality review and queues work on these platforms.

The pay block at the top of this page is computed from the open listings below it: typical (median) hourly rate, the range most roles pay, and how many posted rates it is based on, updated on the date shown. The method is on the pay methodology page.

Listed rates on this page currently run $11 to $23/hr for the middle half of roles.

Fewer listings here state a rate than on the other hubs, and some pay per audio minute or per task rather than per hour; those are excluded from the hourly figure, so read the listing for its own terms. The notes beside this section say when our own estimates or one platform dominate the numbers. Speech AI evaluation and linguistics roles sit at the top of the range; verbatim transcription in a widely spoken language at the bottom.

1. Search by your language first. Use the platform and search filters above the list, or the language pages, and apply only where the dialect matches yours.
2. Prepare your sample the strict way. Transcribe one minute of a podcast in your language verbatim, with speaker labels, timestamps and event tags. It is your practice for the test and your portfolio piece.
3. Apply to two or three platforms. Transcription projects run for weeks, then stop; the next one on the same platform may be in a different language.
4. Check the pay unit before accepting. Per audio minute, per task and per hour are all used here, and a per-minute rate is only good if your speed is.
5. Set the alert on this page so a new project in your language reaches you the day it opens.

Related reading: voice and audio AI training jobs covers the recording side, and one worker's voice recording experience shows what a speech data project is like from the inside. Adjacent kinds of work: RLHF evaluation grades text, red teaming tests refusals, and agentic tasks cover agents that act.