What is In-the-Wild Speech?
Speech collected under fully natural conditions (no studio, no topic constraints), as close to real usage as collection gets.
In-the-wild speech is audio captured in uncontrolled conditions: real environments, real devices, real background, with the recording made as the situation happened rather than staged for the collection. It is the least controlled category and the most representative of deployment.
The defining property is that the variation is not designed. Every other category chooses its conditions; this one records what occurs, which means the corpus is more realistic and harder to characterize. The characterization moves entirely into the metadata.
The uses are robustness evaluation and hard-case training. In-the-wild data is where a system that performs well on curated benchmarks is found to fail, and it is the material that shows which conditions cause the failure.
What buyers get wrong
Real and usable pull against each other. The more natural the audio, the harder it is to annotate. The balance has to be designed into the collection plan, not patched in afterward.
The specification detail that decides everything
Recording device (phone, headset, professional mic) and collection environment must be logged per file.
This is the line item that most often separates a dataset that works from one that gets re-ordered. It belongs in the specification before production starts, not in the delivery review.
How we source it
We do not hold inventory in this category. A requirement comes in, we match it to a producer who can meet the specification, and we agree the quality bar with both sides before recording begins. A pilot batch goes first so misalignments surface early.
Scenarios this covers
Buyers rarely ask for a category by name. They describe a situation, and the situation is what tells us whether this category is the right one. If yours is in the list, name it in the brief and it goes straight into the specification. If it is not, describe it anyway, because it usually means the requirement sits in a neighbouring category.
What good In-the-Wild Speech looks like
In-the-wild data is judged on its metadata, because the conditions cannot be controlled and therefore have to be recorded.
-
Per-file condition documentation
Device, environment, background sound and distance, recorded at capture rather than reconstructed later.
-
Genuine capture rather than reconstruction
Audio recorded in the situation rather than staged to resemble it, because the difference shows in exactly the details that matter.
-
A spread that is reported rather than averaged
The corpus covers a range of conditions, and the distribution is delivered so a buyer can select or stratify.
-
Consent that covers the capture context
Recording in public or shared spaces raises questions a studio session does not, and the consent has to address them before capture begins.
-
Hard cases retained rather than filtered
The difficult files are the value, and a corpus filtered for quality has removed the reason it was collected.
Public datasets in this category
These are openly documented resources, not products we sell. Each one covers part of the category, and knowing where each stops is usually what decides whether a project needs a collection of its own.
-
YODAS
A very large collection of subtitled speech from public video, covering many languages and a wide range of recording conditions. Caption quality is inherited from the platform, so it varies.
-
MSP-Podcast
Podcast and broadcast speech with paralinguistic annotation, recorded in real conditions rather than studios. Realistic and small.
-
Common Voice spontaneous subsets
Most of Common Voice is read speech, and the platform has run collection efforts for spontaneous and conversational material. Those subsets are a small fraction of the total.
-
Why captions are the bottleneck
Corpora harvested from public video are only as good as the captions they were aligned to. Where captions are auto-generated, the transcript errors are systematic and they propagate directly into training.
Questions to ask a supplier
Questions that establish whether the conditions were recorded or merely described after the fact.
- How is the condition of each file recorded, and at what point?
- Was the audio captured in the situation or recreated for the collection?
- What is the condition distribution, and can it be delivered as a table?
- How does consent address capture in public or shared spaces?
- What proportion of files were excluded for quality, and by what criterion?
The vague answer in this category is about the transcript source. A supplier who says the corpus is transcribed without saying whether the transcripts were human or automatic has described a corpus whose ceiling is unknown. In harvested audio, automatic captions are the usual source, and their error rate becomes the error rate of the corpus.
A verifiable answer states the source, the caption provenance, the measured condition distribution and the deduplication method. Ask for the near-duplicate rate, because public video contains reuploads and clips of clips, and a corpus that has not deduplicated reports a size it does not have.
Technical attributes
The items below are the ones a specification can hold a producer to, and the ones that decide whether a delivery is usable for training rather than only present.
| Capture conditions | Recorded per file at the point of capture |
|---|---|
| Device and environment | Documented and varied |
| Condition distribution | Delivered as a table rather than summarized |
| Capture authenticity | In situ, not reconstructed |
| Consent scope | Covers the capture context, including bystanders |
| Delivery format | Audio with a per-file condition table |
The recording condition column is what makes harvested data usable at all, and it is the one most often left coarse. The defining feature of in-the-wild audio is that nobody controlled the capture, so the only way to make it trainable is to describe what happened. Even a rough band per segment is enough to filter on.
The values come from the source platform and the domain. Video from one platform has a characteristic set of devices, distances and environments, and a model trained on it inherits them. Naming the platform in the specification is what lets a buyer judge how far the corpus will transfer.
A mistake here is measured as an unexplainable error rate. A corpus with unknown conditions produces a model whose failures cannot be attributed to anything, and debugging turns into guesswork. The condition labels are cheap to carry and impossible to reconstruct afterwards.
How In-the-Wild Speech is collected
The collection is the situation itself, which makes documentation the whole of the quality control.
- Capture protocol design. What is recorded, on what device, and which condition metadata is captured at the same moment.
- In-situ recording. Audio is captured as the situation occurs, with the device the situation uses, rather than staged in a way that resembles it.
- Condition capture alongside the audio. Device, environment, background and distance are recorded at capture, because reconstructing them afterward produces guesses.
- Consent for the context. Consent addresses recording in shared and public spaces, including how speech from people who are not participants is handled, before any capture happens.
- Minimal filtering with disclosure. Files are excluded only for defined reasons, and the exclusion rate and criteria are delivered with the corpus.
After acquisition, the audio is deduplicated, transcripts are checked for quality and flagged by source, conditions are labelled at least coarsely, and the files are packaged with a manifest stating the measured distributions. The deduplication figure belongs in that manifest, because it changes what the corpus size means.
The pilot for harvested audio is a sample from the platform the specification names, run through deduplication, transcription and condition labelling before the full acquisition. Its purpose is to measure the near-duplicate rate and the caption quality, because those two figures decide what the corpus is actually worth. Both are measured on the sample rather than assumed from the platform.
Acceptance is checked on the deduplicated hours and on the measured condition distribution, against figures stated before acquisition. The specification should name the platform and the caption source in advance. Where the pilot shows the duplicate rate is high, the acquisition set is widened rather than the reported hours adjusted, since a corpus that has not deduplicated reports a size it does not have.
How a delivery goes wrong
The defects below are the ones that recur across this category. Each is decided at collection or annotation time, which is why none of them can be repaired in delivery review.
-
Recording conditions are unknown
The defining feature of in-the-wild data is that nobody controlled the capture. If device, distance, and environment are not documented at least at a coarse level, the corpus cannot be stratified and evaluation cannot be explained.
-
Transcript quality varies within one delivery
Harvested corpora mix human and machine transcripts. A per-segment confidence or a source flag is what allows a buyer to filter, and it is usually absent.
-
Duplicates and near-duplicates
Public video contains reuploads and clips of clips. Without deduplication the corpus size is inflated and the same content appears in both training and evaluation splits.
Why this matters to buyers
The gap between benchmark performance and field performance is the central problem in deployed speech systems, and it is measured by in-the-wild data. A model evaluated only on curated sets has been evaluated on a distribution it will not meet.
The second driver is diagnosis. When a system fails in the field, the useful question is which condition caused it, and answering that requires data where conditions are recorded per file rather than inferred from an aggregate.
A model trained on harvested audio without condition labels fails in ways nobody can explain. The failures are real and the causes are invisible, so the team spends its time on model changes that cannot help. The corpus is usually replaced rather than repaired, which means the collection budget gets spent twice.
The second cost is scale illusion. A corpus that reports a hundred thousand hours and has not been deduplicated may contain a fraction of that in distinct content, and the evaluation sets leak into the training sets. The reported quality is then inflated in a way that shows up only on genuinely new data.
A harvested corpus with a measured duplicate rate and condition labels gives a model whose failures can be attributed. The team can filter by condition when a deployment differs from the source platform, and the effective size is known rather than estimated from a raw figure.
The second gain is that the corpus can be extended across platforms. Each platform has its own devices, distances and environments, and a delivery that documents the condition distribution can take a second source without the two being conflated. That record is what turns a one-off harvest into a growing corpus.
Specification at a glance
| Category | In-the-Wild Speech |
|---|---|
| Group | Speech data |
| Critical spec | Recording device (phone, headset, professional mic) and collection environment must be logged per file. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
In-the-Wild Speech across languages
This category is available in every language we cover. The difficulty changes with the language, which is why each language page documents its own constraints.
Related categories
-
Call Center Speech
Customer-service calls, simulated rather than intercepted, capturing both the agent and the caller side, used for call center QC, intent recognition, and dialogue systems.
-
Conversational Speech
Speech from natural conversation between two or more people, on open or semi-structured topics, used for conversational AI, voice assistants, and small talk models.
-
Singing Voice
Vocal recordings with melody, in both a cappella and accompanied form, used for singing voice synthesis, music information retrieval, and lyric alignment.
-
Speech Commands
Targeted recordings of short command words or phrases, usually with many speakers reading each entry several times over, used for wake words and on-device command recognition.
-
Read Speech
Recordings of speakers reading specified text, with clear pronunciation and known text, the base material for TTS and ASR.
-
Podcast Speech
Long-form podcast and interview audio, either solo monologue or two-person conversation, used for long-form speech recognition and speaker modeling.
Questions about In-the-Wild Speech
Why not filter out the difficult files?
Because the difficult files are the product. A corpus filtered for quality has removed the conditions the buyer is trying to train for, and the result is another clean corpus with a misleading name.
How can conditions be recorded reliably?
At capture, by the person making the recording, using a fixed metadata form. Reconstructing conditions afterward produces descriptions that vary by whoever writes them and cannot be compared across files.
What about bystanders in public recordings?
Consent has to address the capture context before recording begins, including how speech from people who are not participants is handled. This is designed into the protocol rather than resolved afterward.
Is in-the-wild data a replacement for controlled corpora?
No, it is the complement. Controlled corpora train a model under known conditions and give clean evaluation; in-the-wild data shows what happens outside them. Most programmes need both.
Do you have In-the-Wild Speech available now?
No. We do not hold inventory. Everything is produced against a specification. That means the timeline starts when the spec is agreed rather than immediately, and it also means the data matches your requirement instead of approximately matching something already built.
Can this be combined with other categories in one delivery?
Yes, and it is common. A single specification can cover several categories across several languages, with one pilot batch and one quality bar. It usually reduces total cost because contributor recruitment and project setup are shared.
What does the consent documentation cover?
Signed contributor consent stating the intended use, the collection methodology, and the annotation guideline the data was produced under. For projects involving EU data subjects the chain is built to satisfy GDPR, including the transfer mechanism. The terms for moving the data across a border.
Request In-the-Wild Speech
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.