Custom audiovisual collection
Synchronized audiovisual speech data collection with use-specific consent
Design a synchronized audio-visual speech dataset for visual speech recognition, lip alignment, audiovisual agents, or digital-human research, with the target model use stated in the participant release.
Example availability: Tell us the model use and capture conditions. We will check whether a permissioned representative example can be shared; otherwise, we can scope a small pilot.
Collection specification
- Capture
- Synchronized face video and multichannel audio, 1080p to 4K at 25 to 60 fps, sync tolerance agreed per project
- Conditions
- Cameras, lighting, poses, distance, and background noise varied to a written matrix
- Coverage
- Languages, speakers, demographics, and speaking style, scripted or spontaneous
- Metadata
- Speaker, session, camera, lighting, prompt, timing, and approved-use fields per recording
- Rights
- Documented likeness, biometric, and AI-training consent per participant, reviewed before collection
- Status
- Custom collection; catalogue rights are never assumed to cover a new audiovisual use
- Pricing
- Custom, scoped to your conditions
Treat synchronization and permission as product specifications
A video that happens to contain speech is not automatically an audio-visual speech dataset. Frame timing, audio alignment, mouth visibility, capture consistency, metadata, and the intended model use all decide whether the recording can be trained on responsibly.
Face video is biometric-adjacent data, and several of the model uses on this page push it into special-category territory under the GDPR. So every participant signs likeness, biometric, and AI-training language that names the intended use before a camera rolls, which is the documented consent chain scraped footage and public research corpora cannot offer. Existing recordings are never represented as suitable for avatar, lip-reading, or biometric work without a separate rights review.
What audio-visual speech models need from the capture
Lip movement is fast. The visemes that separate similar phonemes resolve within a frame or two, so frame rate and audio-video sync tolerance are set against the phonetic detail the model must recover rather than left at camera defaults. Mouth visibility is specified the same way: head-pose limits, occlusion rules, and lighting that keeps the lower face readable in every condition the model will serve.
Audio-visual speech recognition earns its advantage where audio degrades, and a corpus recorded only in quiet rooms trains that advantage away. Collections for AV-ASR robustness therefore mix conditions to a written matrix: clean rooms, background noise, overlapping speakers, and varied microphone distance, with the visual channel kept clean enough to carry what the acoustic channel loses.
The model job sets the specification
Visual speech recognition weights mouth-region resolution and viseme-balanced prompts. Lip alignment and dubbing weight frame-accurate timestamps and phoneme-level transcripts. Audiovisual agents and digital humans add conversational dynamics: turn-taking, listener behavior, and facial expression recorded across full dialogue sessions rather than isolated read sentences.
Naming the job first keeps the scope honest, because each one changes what counts as a usable recording. Talking-head avatar training has a dedicated offer with its own capture and rights profile, as does broader video data collection beyond speech; this page covers the synchronized audio-visual core they share.
Agree the audiovisual pilot before scaling capture
Use this illustrative checklist to discuss a first batch. It is a scoping template, not a customer recording or a promise that a particular sample is available.
- Capture matrix
- Specify speaker and language quotas, prompt types, camera angles, lighting, microphone placement, and noise conditions.
- Sync acceptance
- Agree the audio-to-video offset limit, how it is measured, and which clips are rejected or re-recorded.
- Manifest fields
- Define clip, participant, session, camera, audio, prompt, timing, consent-record and approved-use references.
- Pilot decision
- Review a rights-cleared first batch against visibility, timing, metadata and permission checks before approving production volume.
The proposal should state the batch size, review method, rework rules, rights boundary and deliverable schema; actual examples are shared only when permission permits.
This offer is not the right fit when
- You want to scrape public faces or reuse footage without participant permission.
- You need existing catalogue video before its synchronization and rights have been reviewed.
- Your target model use cannot be described to participants.
Check example availability
- Your details
- Your project
- Verify
Three short steps. The team checks whether a permissioned example matches your use and can be shared, then follows up with the available options.
Would rather talk it through first? Scope an audiovisual pilot
Prepare the brief
Bring a specification your vendors can price consistently
Use the free worksheet to define coverage, rights, delivery fields, held-out rules, and acceptance tests before requesting a collection plan.
Buyer documentation
Specimen data card and provenance structure, evaluation brief template, and acceptance and held-out questions.
Related services
Guides for this decision
Need a scoped collection plan? Share the visual speech task, speaker matrix, camera and microphone setup, sync requirement, rights scope, and first-batch size you know. The quote form opens with audiovisual fields selected.