Audio annotation is the step that turns a folder of recordings into something a model can learn from. The recordings are cheap to gather. The labels, the transcripts, the speaker turns, the timestamps, the event tags, are where most of the cost and most of the risk sit. Choosing an audio annotation vendor is really a decision about who controls the quality of your training data.
This guide is written for the people who have to sign off on that decision: ML leads, data managers, and founders hiring an audio annotation service for the first time. It covers what the different services actually deliver, how they are priced, how to judge quality before you commit, and when building in-house makes more sense than buying.
What Audio Annotation Services Deliver
"Annotation" is a catch-all word. When you brief a vendor, be specific about which of these you need, because each one changes the price and the turnaround.
Verbatim transcription. Every word that was said, written down, usually with a decision about how to handle filler words, false starts, and repetitions. Verbatim keeps them. Clean or intelligent transcription removes them. Voice AI training almost always wants verbatim, because the model has to learn from disfluent, messy human speech, not a tidied-up version of it.
Speaker diarization. Marking who spoke when, so a two-person call becomes Speaker A and Speaker B with turn boundaries. This is essential for call-center data, interviews, and any multi-party audio. Diarization is harder than it sounds when speakers overlap or sound similar, and it is a common place for quality to slip.
Timestamping. Aligning text to audio, either at the segment level (this sentence starts at 00:14) or the word level. Word-level alignment is more expensive and is what you need for text-to-speech and for training forced-alignment tools.
Event and non-speech tagging. Labels for laughter, coughing, music, silence, background noise, or channel artifacts. Wake-word and voice-activity models depend on these tags being consistent.
Semantic labels. Intent, sentiment, topic, or entity tags layered on top of the transcript. This is where annotation shades into natural language work, and it needs a clear label taxonomy or agreement collapses.
Phonetic and prosodic markup. Phoneme boundaries, stress, and intonation for high-end text-to-speech. This is specialist work and a small subset of vendors do it well.
Ask for a sample deliverable in the exact format you will consume, before any money changes hands. A vendor that cannot produce a clean sample JSON on request is telling you something.
How Annotation Is Priced
Pricing is usually per audio minute, per audio hour, per item or per batch, and it scales with how many of the labels above you stack on top of each other. A reliable quote needs a sample of your audio and a written spec, because the same label costs very different amounts on clean studio speech and on noisy calls. We price our own annotation per audio hour, item or batch, and quote after reviewing a sample.
| Annotation type | What drives the cost |
|---|---|
| Verbatim transcription | Accents, background noise, and the conventions for fillers, numbers and unclear speech |
| Transcription plus diarization | How many speakers there are and how often they overlap |
| Word-level timestamps | How tight the boundary tolerance is; needed for TTS and alignment |
| Event and non-speech tagging | Added on top of transcription; depends on tag granularity |
| Intent or sentiment labels | How large and clear the taxonomy is; needs a fixed tag set |
| Phonetic and prosodic markup | Specialist linguists, so it depends heavily on the language |
Three things move the price more than anything else. Language is the first: a widely spoken language with a deep annotator pool is cheap, and a low-resource language costs more because there are fewer qualified people. Audio quality is the second: noisy, far-field, or overlapping speech takes longer to label and needs more review. Domain is the third: medical, legal, and financial audio require annotators who understand the terminology, and that expertise carries a premium.
Be suspicious of a quote that is far below the others you get for the same spec. Cheap per-minute pricing usually means single-pass labeling with no independent review, and you pay for it later in rework and in models that learn from bad labels.
How to Judge Quality Before You Commit
The single most useful thing you can do is run a paid pilot on a small, representative batch before signing a large contract. A few hours of your real audio, labeled to your real spec, tells you more than any sales deck. Here is what to measure.
Inter-annotator agreement. Have two annotators label the same files independently and compare. For categorical labels, compute Cohen's kappa or Krippendorff's alpha. For transcription, compute word error rate between the two versions. If two of the vendor's own people disagree badly, your training data will be inconsistent no matter how the numbers look on average.
Accuracy against a gold set. Build a small reference set yourself, or with an expert, and score the vendor's output against it. For transcription, this is a straightforward word error rate. Decide the threshold up front. For clean English verbatim, expect under 5 percent WER from a good vendor. For noisy or conversational audio, 8 to 12 percent may be realistic, and you should agree on that number before work starts.
Guideline adherence. Give the vendor a written annotation guide and check whether the pilot follows it, especially the edge cases: how they handle unintelligible speech, overlapping talk, numbers, and non-standard spellings. Consistent handling of edge cases separates a mature annotation operation from a cheap one.
Turnaround and communication. Note how long the pilot took and how the vendor handled your questions. Annotation projects live and die on the feedback loop. A vendor who asks clarifying questions early is worth more than one who silently guesses and delivers fast.
The Questions to Ask a Vendor
Bring these to the first serious call. The answers separate real annotation operations from resellers.
- Who does the labeling, and are they employees or an anonymous crowd? Both models work, but they have different quality profiles and different privacy implications.
- What is your review process? A single pass with no review is a red flag. Look for at least a second-pass review or a sampled audit on every batch.
- How do you measure quality, and will you report agreement scores and error rates with each delivery?
- How do you handle data security and privacy? Where is the audio stored, who can access it, and can annotators download it? This matters enormously if your recordings contain personal information.
- Can you handle my languages and my domain, and can you show relevant past work?
- What happens when I reject a batch? Rework terms should be in the contract, not improvised later.
When to Build In-House Instead
Buying is not always right. Build an internal annotation capability when annotation is genuinely core to your product and improves as a moat, when the work needs domain expertise that is slow and expensive to transfer to an outside team, or when your data is too sensitive to leave your environment at all. Medical audio under strict regulation often falls into that last category.
The common middle path is a hybrid. Keep a small internal team that owns the annotation guidelines, designs the label taxonomy, and audits quality, and use a vendor for raw volume. The internal team defines what "correct" means and polices it. The vendor scales it. This keeps control where it matters while still getting you throughput.
A Note on Buying Data That Is Already Labeled
Sometimes the fastest path is not to annotate at all, but to license data that is already collected, transcribed, and labeled to a known standard. Off-the-shelf speech datasets come with the annotation baked in and quality already measured, which removes the vendor-selection problem entirely for common use cases. It will not fit every need, especially narrow domains where audio has to be collected and labeled to your spec, but for general conversational speech recognition or text-to-speech it is often cheaper and faster than commissioning annotation from scratch.
If you do commission annotation, treat the vendor relationship the way you would treat any critical supplier. Pilot before you scale, measure quality with numbers rather than impressions, and keep ownership of the definition of quality inside your own team. The annotation is the part of your dataset the model actually learns from. It deserves that scrutiny.
If you are scoping a project now, our audio annotation services page lists the layers, formats and QA terms we agree before production.
