Transcription, Audio & Speech

Looking for transcription jobs? The roles at the top of this page are remote transcription and audio jobs in AI training: platforms pay people to transcribe, correct and label recorded speech so speech-recognition and voice models can learn from it. The core skill is the same as in traditional transcription (listening closely and typing exactly what was said, to a style guide), with two differences worth knowing before you apply. AI-training transcription is usually paid by the hour or by the task rather than per audio minute, which spares beginners the per-minute trap explained below. And most of these roles are language-specific, so native fluency in an in-demand language (such as French, Spanish, Portuguese, German, Korean, Chinese or Hindi) is the fastest way in and up. The list is refreshed from each platform's own job board every day.

Audio, speech and transcription work in AI training means listening to recorded human voices and turning them into accurate, machine-usable data: typing out what was said, marking who spoke when, tagging background noise, labeling emotion and intent, or recording your own voice to teach a model how real people talk. It is one of the most accessible ways to get paid to train AI, and the honest headline is this: the transcription, audio and speech roles open on this site today typically pay Unavailable to Unavailable an hour, with a median of Unavailable. Entry-level generalist work on the big crowdsourcing platforms ( Appen, Clickworker, TELUS) pays below that band, per-audio-minute transcription pays whatever your speed makes it, and specialist, multilingual or expert-tier work sits at the top of it. Which end of that range you land at depends far more on your language, your speed, and the specific project than on the platform's name.

This guide is written for a real person deciding whether this work is worth their time. It covers exactly what the work is day to day, the platforms that hire for it and what each actually pays (by platform, by seniority, and by country), the skills and gear you need, how to pass the unpaid assessments, realistic earnings math (including the trap that sinks most beginners), taxes and payment logistics, the pros and cons, the scams that prey on people looking for this exact work, and where the field is heading in 2026.

What Transcription, Audio and Speech Work Actually Is

The category sounds narrow but covers a spread of tasks, and the task type is the single biggest driver of both pay and difficulty. Speech data work generally falls into three buckets: producing speech (you record your voice), transcribing speech (you type what you hear), and annotating speech (you add structured labels on top of audio). Most platforms mix all three.

Transcription

Transcription is the most common speech annotation technique: turning spoken language into accurate text ( Toloka). For AI training there are two flavors that pay differently and demand different discipline:

• Verbatim transcription captures every "um," false start, stutter and repetition, producing high-fidelity training data for speech models. It is slower and more exacting.
• Clean (or "clean read") transcription removes filler words and false starts for readable text. It is faster but requires judgment about what to cut.

You will also see timestamping, where you map words or segments to exact time positions in the audio. Timestamps record when each word or phrase begins and ends, which meeting transcription, subtitles and speaker-diarization training need ( Digital Divide Data).

Speech and audio annotation

Annotation is where you add labels a machine can learn from. The main types you will encounter ( Toloka, Digital Divide Data):

• Speaker diarization: assigning each segment to the correct speaker ID and marking turn-taking boundaries when one speaker stops and another starts. Essential for meeting transcription and multi-party voice assistants.
• Intent and entity labeling: categorizing commands so an audio cue like "play music" triggers the correct action. This is the backbone of virtual assistants (Alexa, Siri, Google Assistant).
• Prosody and phonetic marking: labeling stress, pitch and pronunciation, often for text-to-speech models.
• Emotion and sentiment labeling: tagging whether a speaker sounds angry, happy, frustrated or neutral.
• Noise and acoustic-event tagging: marking background conditions such as traffic, music, keyboard clicks or microphone interference so models learn to hear through noise.
• Language ID and accent handling: identifying which language or dialect is spoken, including code-switching within one clip.

Speech data collection (recording your own voice)

The third bucket is the easiest to start and usually the lowest paid per task. Platforms pay you to record utterances: single words, short scripted sentences, smartphone commands like "stop the music," or short natural conversations. Clickworker's audio program, for example, has you record single words or short sentences on your smartphone to train speech-recognition tools. DataForce and Appen run large in-app or remote voice-collection drives, sometimes for specific accents, ages or languages they are short on.

The takeaway: "transcription" is a skill you get faster at and get paid more for over time; "recording utterances" is quick-cash micro-work with a low ceiling; and "annotation" sits in between and rewards people who follow guidelines precisely.

The Platforms That Hire (and What They Really Pay)

There is no single rate for this work. The same platform pays a US worker far more than a worker in the Philippines for the identical task, and a specialist project can pay several times a generalist one. Below is what each platform pays: its own published figures with the day they were read, or the rate card this site keeps for it, which opens the platform's pay page and the day the card was checked. Treat every number as a realistic range, not a guarantee, because task availability, project cycle and your own speed move the effective rate more than the advertised one.

Pay by platform

A few things jump out. First, the "big four crowdsourcing" platforms named as top hirers for this niche (Appen, Sigma AI, DataForce, TELUS, plus Clickworker) cluster at the lower end for pure audio work, as their rate cards above show. Second, the newer "train the LLM" marketplaces ( Outlier, DataAnnotation) pay noticeably more but gate the best work behind credentials and geography. Third, Sigma AI is the cautionary tale of the group: it is a real company, but transcriber reviews rate its compensation poorly, with contractors reporting that a short recording can take many times its length to transcribe to standard while they are only paid for the audio's length ( Glassdoor).

Pay by seniority and specialization

Within any single platform, what you can do moves your rate more than which platform you are on.

Outlier illustrates this: a pay analysis by Breaking Even, updated 14 Sep 2026, has workers at roughly $12 to $18 per hour in weeks one and two while learning, $18 to $25 by weeks three to six, and $22 to $35 from month three on a stable project, with expert-tier legal, medical and scientific work at $35 to $60. The lesson: pick up a rare or professional language, or a professional domain, and you leave the generalist floor behind.

Pay by country

Geography is the harshest variable in this work, and platforms are explicit that they pay by locale. For the same task:

TELUS's rate card, checked 31 Oct 2026, reads $18–$50/hr experts; linguists and evaluators $3.50–$25/hr, and where you sit in that range depends heavily on your country and language. The practical implication for workers outside the top markets: your leverage is language rarity and speed, because the base geographic rate is fixed and low. Country restrictions are also the single biggest filter on access. DataAnnotation publishes no country list, and worker reports point to the US, Canada, the UK, Ireland, Australia and New Zealand, while Appen, TELUS and Outlier offer much wider global access at lower rates.

The Per-Audio-Minute Trap (the Math That Sinks Beginners)

This is the most important section in the guide, and the one thing most people wish they had understood before starting. Many transcription platforms (Rev, TranscribeMe, Sigma) pay per audio minute, not per minute worked. That distinction is where advertised pay and real pay diverge.

The industry standard is that it takes an experienced transcriptionist 4 to 6 hours to transcribe one hour of audio, a 4:1 to 6:1 ratio of work time to audio length ( GMR Transcription, 5 Feb 2025). Beginners and poor audio can push that to 6:1 or worse. So the arithmetic works like this:

• Rev pays $0.40 to $1.10 per audio minute for transcription ( Rev, read 29 Oct 2026). At $0.50/min and a beginner's 4:1 ratio, one hour of audio pays $30 for four to six hours of work, roughly $5 to $7.50 per real hour.
• TranscribeMe's $15 per audio hour for beginning general transcribers ( its help page, read 29 Oct 2026), at 4:1 to 6:1, works out to $2.50 to $3.75 per real hour for beginners.
• Sigma pays per minute of audio; contractors report a short clip taking many times its length, which is why its compensation is rated so poorly ( Glassdoor).

The single lever that fixes this is your ratio. A Rev review updated August 2024 calls a 3:1 ratio (three minutes of work per audio minute) "good" ( The Work at Home Woman). Get to 3:1 or better, and at Rev's upper rates of $0.80 to $1.10 per audio minute your effective rate climbs to $16 to $22 per hour. Speed comes from a good foot pedal, mastering keyboard shortcuts, expander/autotext software, and picking clean-audio projects. On per-audio-minute platforms, your typing speed and audio quality are your wage.

Beyond the ratio, two more things quietly erode the headline rate:

• Unpaid quality audits and rejections. On Appen and similar platforms, rejected work and failed quality audits are not paid, so a high rejection rate silently cuts your effective wage ( EarnifyHub).
• Work availability, not rate, is the real ceiling. Reviewers of both Appen and TELUS repeatedly say the bigger issue is not the hourly rate but the number of hours available: projects pause, reduce hours without notice, or get cancelled ( Appen review). You can qualify for well-paid work and still earn little if there is no work in the queue.

Skills, Qualifications and Tools You Need

The good news: for most entry-level audio and transcription work, you do not need a degree. You need language fluency, a good ear, patience, and attention to detail. The platforms gate on demonstrated skill (via an assessment), not credentials.

Skills that matter

• Native or near-native fluency in the target language, including slang, idiom and accents. This is non-negotiable and the hardest thing to fake in an assessment.
• Typing speed and accuracy. For per-minute transcription, a fast, accurate typing speed changes your economics. Every extra word per minute is money.
• A precise ear. You must distinguish similar-sounding words, catch mumbled speech, and handle overlapping speakers.
• Discipline with guidelines. Annotation projects live or die on consistency. Most assessments test your ability to follow detailed guidelines exactly, not your intelligence ( TELUS review).
• Grammar, punctuation and formatting. Clean transcripts require judgment about spelling of names, numbers, and punctuation conventions per the project style guide.

Domain skills that unlock premium pay

• A second in-demand language. Bilingual contributors (Spanish, Arabic, Mandarin, Portuguese, French, Nordic languages) access projects at premium rates ( Outlier review).
• Professional domain knowledge. Medical or legal transcription commands far higher rates because the vocabulary is specialized and errors are costly.

Gear and software

• A quiet room and good closed-back headphones. You cannot transcribe what you cannot hear.
• A reliable computer and stable internet. Most annotation tools are browser-based.
• A smartphone with a decent built-in mic for voice-collection tasks ( Clickworker).
• A USB foot pedal (for example an Infinity pedal) to play, pause and rewind without leaving the keyboard. This alone can noticeably improve your ratio.
• Text-expander software to auto-fill repeated phrases, speaker labels and formatting.
• A PayPal account, which is how most of these platforms pay ( voiceworker.io).

How to Get Started and Pass the Assessments

The path onto every legitimate platform is the same shape: create a profile, declare your languages and locale, and pass an unpaid skills assessment before you can see paid projects (DataForce is the exception: it lists its projects openly on its Projects page, with pay given in each job announcement, read 10 November 2026) ( voiceworker.io, TELUS review).

Step by step

1. Apply directly on the platform's own website. Never through a Discord DM, Telegram group or a link someone sent you (see the scams section). Go to appen.com, telusinternational.ai, dataforcecommunity.transperfect.com, clickworker.com, rev.com/freelancers, or the official Sigma careers site.
2. Build a complete, honest profile. List every language you genuinely speak, your exact locale, your equipment and any domain expertise. Projects are matched to your profile, so an incomplete profile means fewer invitations.
3. Take the qualification assessment. These are unpaid and test attention to detail and your ability to follow detailed guidelines. Read the guidelines twice before you start: the test is of following them.
4. Start with the simplest projects to build a quality history. Your rejection rate and quality score follow you and gate access to better-paid work.
5. Diversify across two or three platforms. Because work availability is the real constraint, having accounts on multiple platforms smooths out the gaps when one project pauses.

Passing tips that actually move the needle

• On transcription assessments, follow the exact style guide (spelling out numbers vs digits, how to mark inaudible sections, speaker-label format). Grading is mechanical.
• On annotation assessments, when in doubt, match the example annotations the guidelines provide rather than your own instinct.
• If an assessment or "training" is vague about time and pay, ask a polite question about expectations. A legitimate employer answers; a scam ghosts you or gets evasive ( FTC).

Realistic Earnings: What You Can Actually Make

Setting the hype aside, here is a grounded picture across three profiles.

• The casual side-hustler (entry, English, few hours/week). On Clickworker utterance tasks or entry Appen audio, expect the low end of their rate cards (Clickworker: Per task, cents to tens of euros; no hourly rate; Appen: $2.40–$100/hr; CrowdGen experts up to $450/hr). This is coffee money, useful for filling dead time, not a paycheck.
• The committed transcriber (fast, quality history, US/UK). With a 3:1 ratio on Rev's higher per-minute rates ($16 to $22 per real hour, as worked out above), plus steady TELUS or DataForce projects, a reliable part-time income is achievable when work is available. The constraint is queue depth, not rate.
• The specialist (multilingual or professional domain). Language-specific and expert projects sit at the top of the audio band (up to Unavailable in the roles open on this site today) and in the multilingual band of Unavailable to Unavailable. This is where the work becomes a genuine income, and it is gated by credentials and geography.

The honest summary: as a primary income, entry-level audio/transcription work is thin unless you are a fast specialist in a strong-paying country. As a flexible supplement you can do from anywhere on your own schedule, it is real, legitimate money, and it is one of the lowest-barrier ways into the broader AI-training economy.

Taxes and Payment Logistics

Almost every platform in this space treats you as an independent contractor (1099 in the US), not an employee. That has direct consequences ( Intuit/TurboTax, IRS):

• No taxes are withheld. You receive the full amount and owe tax later. Set aside a portion of every payout.
• You owe self-employment tax. In the US that is 15.3% (12.4% Social Security plus 2.9% Medicare) on top of income tax, per the IRS (read 29 Oct 2026), filed via quarterly estimated payments on Form 1040-ES ( ADP).
• You may get a 1099-K from PayPal if your payouts cross the threshold, but you must report the income and pay tax on it even if no form arrives ( TurboTax).
• Payment timing varies. DataForce, for instance, pays remote work only after a project's quality check clears and on-site studies right after the session, by PayPal, gift card or other methods by location (its FAQ, read 6 Nov 2026), and Appen pays monthly. Cash flow is not instant.
• You can deduct legitimate expenses. Your foot pedal, headphones, a share of internet, and transcription software are deductible business expenses that reduce taxable income. Keep records.

Outside the US, similar self-employment rules generally apply in your own country. The universal principle: this is self-employment income, so track it and reserve for tax from day one.

Pros and Cons: An Honest Ledger

Pros

• Genuinely low barrier to entry. No degree, minimal gear, just fluency and an assessment.
• Real flexibility. Work when and where you want, in the gaps of your day.
• A foot in the door of the AI economy. Skills and quality history here transfer to higher-paying evaluation and specialist work.
• Legitimate, established payers exist. Appen, TELUS and DataForce are real companies with real clients and traceable PayPal payouts.
• Language and domain skills are directly rewarded with meaningfully higher pay.

Cons

• Low floor for generalists. Entry English audio transcription at Appen is listed at about $8 to $12 an hour in the US and UK and less elsewhere ( Appen review), and per-audio-minute pay can fall far lower per real hour.
• The per-audio-minute trap quietly halves or worse the wage of slow transcribers.
• Unpredictable work availability. Projects pause, cut hours or cancel with no notice, the number-one complaint across platforms.
• Unpaid assessments and rejected work. Time invested in qualification and audits is not always compensated.
• Geographic pay gaps put workers outside the top-paying markets at a structural disadvantage.
• A minority of real companies pay poorly or slowly (Sigma AI's contractor reviews are a warning).

Red Flags and Scams to Avoid

Because "get paid to train AI from home" is a magnet phrase, scammers target exactly the people searching for this work. The legitimate platforms have a consistent shape; the scams break that shape. Here is how to tell them apart ( FTC, DataAnnotation):

• Never pay to work. No legitimate platform charges a fee to apply, "get certified," or "unlock" earnings. Any request to "top up" your account with your own money is a scam, full stop.
• Legitimate recruiters do not DM you on Discord, Telegram or WhatsApp. Real platforms make you apply through their site. A recruiter emailing from a @gmail.com or @hotmail.com address, or pushing to move the chat to Telegram, is fake.
• Check the exact domain. Look-alike addresses (dataannotation.tech vs dataannotation-tech.com, say) are an easy trick. Type the URL yourself; do not click a sent link, and apply only through the platform's own sign-up, as DataAnnotation's trust page asks.
• Beware PII harvesting. Fake "assessments" that ask for your full name, date of birth, bank details or ID before any work are often identity-theft fronts. A real assessment tests skill, not your Social Security number.
• Watch for "too good, no experience needed, guaranteed high pay." Real ranges are modest and gated behind assessments.
• Sudden bans before payout are a documented scam pattern: work collected, account deactivated, no payment. Prefer platforms with public payout records and PayPal/traceable payment.

Legitimate platforms gate projects behind skills assessments, publish concrete pay ranges, and pay through traceable services. If a "job" is missing all three, walk away.

The 2026 Outlook: Is This Work Going Away?

The demand side is booming and the nature of the work is shifting, both at once. The AI annotation market is projected to grow from $2.51 billion in 2026 to $7.32 billion by 2030, a compound annual growth rate of 30.7 percent ( Research and Markets, February 2026), and Grand View Research puts the data annotation tools market at $1.7 billion in 2025 heading to $10.3 billion by 2033 ( Grand View Research, read 29 Oct 2026). Every voice assistant, meeting-transcription tool, in-car system and multilingual model needs fresh, accurately labeled speech data. That demand is not shrinking.

What is changing is what humans do inside it. The best speech-to-text models reach about 95 to 98 percent word accuracy on clean audio but about 70 to 90 percent on accented and noisy audio ( AssemblyAI, 26 Oct 2026), so a hybrid workflow uses automated drafts, then human review concentrated on the tricky, high-value cases ( Annotera). Humans are moving from typing every word to validating and supervising machine output: the quality guardians of AI ( Annotera).

The market is bifurcating as a result. Generalist, English, clean-audio transcription faces the most automation pressure and the most downward rate pressure, because a machine can draft it and a human only cleans it. Specialists face growing demand: rare languages, code-switching, heavy accents, noisy real-world audio, emotion and intent, and professional domains are exactly the cases machines still get wrong and humans are paid a premium to fix ( The AI Rankings).

The strategic read for anyone starting in 2026: enter through accessible transcription and recording work to build a quality history, then move deliberately toward the parts machines cannot do (a second language, a professional domain, verification and quality-review roles). That is where the pay is climbing, and where the work is durable.

Bottom Line

Audio, speech and transcription work in AI training is legitimate, flexible, and genuinely open to people without degrees: the roles open on this site today typically pay Unavailable to Unavailable an hour, per-task and per-minute work often less, and the pay explorer sets that against every other kind of work. Where you land depends almost entirely on your language, your speed, your specialization and your country. The low end is real and worth going in with eyes open: entry English generalist work on the big crowdsourcing platforms pays little (Appen lists audio transcription at about $8 to $12 an hour in the US and UK), and the per-audio-minute pay model punishes slow transcribers. The high end is equally real: multilingual and professional specialists sit at the top of the band, and the demand for exactly that work is projected to grow into a multi-billion-dollar market by 2030. Start on the reputable, traceable platforms (Appen, TELUS, DataForce, Rev, Outlier), never pay to work, guard your ratio and your quality score, set aside money for taxes, and treat the entry work as the on-ramp to the specialist work where the money actually is.

Sources

Every live pay figure above is read from the open roles on this site, and every other pay figure, ratio and market number links to its source inline with the date it was read or published. See the full source list below for the primary platform pages, review sites (Glassdoor, Trustpilot), tax authorities (IRS, TurboTax) and market-research reports used to build this guide.