Short answer. Far more than a per-unit price suggests, and for reasons specific to language. The same hour of audio can cost roughly $0.46 through an automated API or up to around $120 through a professional human service, and low-resource languages carry a premium at every level because qualified annotators are scarce. The bigger economic fact is that costs do not scale linearly with languages: each new language resets a fixed setup cost that volume within a language would otherwise amortise.
Why is multilingual data priced differently from English data?
Because the cost driver is not the task, it is the availability of people who can do it. Annotation in English, Spanish, French or German is cheaper than in low-resource languages where qualified annotators are scarce.
Pricing guides in this sector are consistent on the point: the same task costs more in a low-resource language, and specialist audio work such as low-resource speech commands a significant premium at every level of complexity. Multilingual projects that require cultural competence rather than translation carry a further premium, because the annotator has to understand connotation, regional variation and context-dependent meaning.
Three structural reasons sit behind that.
There is no standing labour pool. For most languages beyond the top twenty, contributors have to be found, screened and trained before any work begins. That is a real cost incurred before a single item is delivered.
Quality requires a second speaker. Verification cannot be outsourced to a cheaper adjacent language. Every hour reviewed needs another person who speaks the same variety.
Scarcity has pricing power. Where only a small number of people can do a task, the rate reflects it, and retention becomes a cost item rather than an HR concern.
The practical implication is that a per-unit rate quoted for English tells you almost nothing about what the same specification will cost in Sylheti or Wolof, and a supplier quoting a flat rate across a diverse language list is either averaging heavily or has not priced the tail.
What does one hour of speech data actually cost?
Anywhere from under a dollar to around $120, depending entirely on whether a person is involved. That range is the single most important number in this field.
The reference points are public. Automated speech-to-text APIs sit at the bottom: roughly $0.15 to $0.21 per audio hour for batch processing at the cheapest tier, around $0.46 for another major provider, about $0.96 for Google Cloud and near $1.44 for AWS Transcribe as of 2026. Professional human transcription runs $1.00 to $3.00 per audio minute, which is $60 to $180 per audio hour, with one widely used service at $1.99 per minute, about $119 per hour.
The gap looks absurd until you look at the labour underneath. A trained transcriptionist takes three to six hours to produce one finished hour of transcript. At modest wages the labour alone is $40 to $80 per audio hour before quality review, project management and margin. The human price is not a markup on the machine price. It is a different activity.
For collected multilingual data the picture is more involved still, because transcription is only one line in the bill. An hour of usable, delivered speech data also carries recruitment and screening, the recording session itself, consent handling and compensation, verification by a second speaker, adjudication of disputes, and the compliance and project overhead that makes the result auditable.
Why don't costs scale linearly with the number of languages?
Because each language carries a fixed setup cost that has nothing to do with volume. Doubling the hours in one language is cheap. Adding a second language is not.
This is the least understood aspect of multilingual budgeting, and the one that produces the most unpleasant surprises.
Fixed per language, regardless of volume: recruitment and screening, guideline translation and adaptation, a pilot batch and its review, dialect and orthography decisions, legal and consent review for the jurisdiction, and building a gold-standard reference set.
Variable per hour, and falling with scale: recording, transcription, verification and delivery.
The consequence is that unit cost drops sharply as volume grows within a language and resets every time a language is added. A programme of 1,000 hours in one language and a programme of 100 hours in each of ten languages have similar headline volumes and very different costs.
Two practical readings follow. First, going deep in fewer languages is usually cheaper per usable hour than going shallow in many, which matters because thin per-language volume also fails to clear the threshold at which data helps a model.
Second, a supplier with existing verified contributors in a region starts from a lower fixed cost than one who begins recruiting after signature, which is a real difference in price rather than a marketing claim.
Where do multilingual budgets actually break?
On rework. The cost of fixing a problem rises by roughly an order of magnitude at each stage it survives, and language projects are unusually good at hiding problems until late.
The blunt version comes from an annotation pricing guide and applies across the field: the real cost is not the per-label price, it is the cost of fixing a model trained on poorly labelled data.
Four failure patterns account for most overruns.
Ambiguous guidelines. An instruction that can be read two ways is invisible at pilot scale and produces a split dataset at production scale. Fixing it after delivery means re-reviewing everything.
Recruitment shortfalls. A language that cannot fill its contributor quota stalls the schedule while every other language runs, and delivery dates are usually set collectively.
Specification mismatch. Recordings that satisfy every automated check and none of the intent. This is the single most expensive category, because the work was done correctly against the wrong understanding.
Late compliance discovery. Finding out after collection that reviewer location or consent scope does not permit the workflow, which can invalidate a batch outright.
None of these are exotic, and all of them are cheapest to prevent at the pilot stage, which is exactly the stage most often cut to save two weeks.
When is automation cheaper, and when is it a false economy?
Automation is cheapest where a machine is competent, which correlates almost exactly with the languages that least need new data.
The honest case for automation is strong in the right conditions. Machine pre-transcription followed by human correction is materially cheaper than transcription from scratch when the automated draft is good, because correcting is faster than typing. For high-resource languages with clean audio, this hybrid is now the default and the economics are genuinely favourable.
The case weakens sharply along three axes.
Language. Automated accuracy falls with resource level, and past a threshold correcting a bad draft takes longer than starting fresh, because the transcriber is fighting the machine's errors as well as the audio.
Conditions. Field recordings with background noise, overlapping speakers and regional accents degrade automated output well below the point where it saves time.
Consequence. Where the data trains a safety-relevant or regulated system, verification cost does not fall just because a draft was generated cheaply.
There is also a subtler trap. Using an English-centric model to pre-label or generate data in a low-resource language imports that model's blind spots, and those errors are expensive precisely because they look plausible. A cheap draft that produces confident, wrong output can cost more than no draft at all.
The workable rule is to let automation do what is measurable and let people do what requires judgement, then price the two separately rather than blending them into a single per-unit rate that hides which is which.
How should a multilingual programme be budgeted?
Per language, with the fixed costs stated separately, the pilot funded properly, and an explicit allowance for attrition.
Six practices make budgets survive contact with delivery.
Budget per language, not per programme. A blended average conceals which languages are subsidising which, and makes it impossible to decide rationally what to cut.
Separate fixed from variable. Show recruitment, guidelines, pilot and legal review as their own lines. This makes the cost of adding a language visible before it is committed.
Fund the pilot. It is the cheapest place to find every problem that would otherwise be found at scale.
Allow for attrition. Recorded hours and delivered hours are not the same number. A budget built on delivered hours without a rejection allowance will overrun.
Price quality explicitly. Verification, adjudication and gold-set maintenance are what make the data usable. A quote that omits them is not cheaper, it is incomplete.
Compare like for like. Two quotes at different prices are usually specifying different things: review coverage, dialect breadth, consent handling and documentation. Ask what each excludes.
This is also where an existing delivery footprint changes the arithmetic. Because Lifewood already has screened and trained speakers working through delivery centres across more than 30 countries, a large share of the fixed cost in a given language has been paid once rather than being rebuilt per project. That is not a quality argument; it is a cost structure one, and it is usually the largest single variable between two otherwise similar quotes.
Key takeaways
- The same hour of audio costs roughly $0.15 to $1.44 through automated APIs and $60 to $180 through professional human transcription.
- A trained transcriptionist takes three to six hours to produce one finished hour of transcript, so labour alone is $40 to $80 per audio hour at modest wages.
- Low-resource languages carry a premium at every complexity level because qualified annotators are scarce.
- Multilingual work requiring cultural competence rather than translation carries a further premium.
- Costs are fixed per language for recruitment, guidelines, pilots, dialect decisions and legal review, and variable per hour for recording, transcription and verification.
- Unit cost falls with volume within a language and resets with every language added, so depth is usually cheaper per usable hour than breadth.
- Budgets break on rework: ambiguous guidelines, recruitment shortfalls, specification mismatch and late compliance discovery.
- The real cost of annotation is not the per-label price but the cost of fixing a model trained on poor labels.
- Machine pre-transcription with human correction is genuinely cheaper for high-resource languages and clean audio, and can cost more than starting fresh below a quality threshold.
- Budget per language, separate fixed from variable, fund the pilot, allow for attrition, and price verification explicitly.
Sources and further reading
- ConvertAudioToText, "Cost of Transcription Per Hour in 2026", on verified API rates, human service rates and the three-to-six-hour transcription ratio
- SpeakWrite, "Transcription Costs in 2026", on human transcription rate ranges
- AssemblyAI, "Speech-to-Text API Pricing", on current per-hour API pricing
- DataVLab, "Data Annotation Pricing 2026", on the low-resource and cultural competence premium
- DataX Power, "Data Annotation Pricing in 2026", on task-level cost drivers across modalities
- Analytics & AI, "How Much Does Data Annotation Cost in 2026?", on per-label ranges and remediation cost
- Lifewood, company overview and delivery network