From a 150+ language corpus, 50+ productionised for in-cabin. Native-speaker, real cabin noise, wake-word and DMS audio included. ADAS perception data on the same engagement.
DPA included · EEA-only processing · Project lead replies within one business day
Evaluation engagements, data-collection programs, and active annotation work across OEM and Tier-1 suppliers.
Cabin noise, accent drift, and Article 10 evidence gaps cost more programs than perception ever does. Four places the data, not the model, is the bottleneck.
In-cabin voice fails
Highway cabins reach ~70 dB SPL at 110 km/h: engine drone, HVAC, road, passenger overlap. Lab-recorded corpora flatter the model in eval and ship false rejects to the customer on day one. The fix is recording in the actual cabin, not cleaning up booth audio after the fact.
Coverage gaps
Nordic dialects, Eastern European accents, and code-switching cabin conversations (driver in one language, passenger replying in another) are the long tail US-domiciled vendors do not cover natively. YPAI captures these in-region with native speakers, not synthetic accent transfer.
EU AI Act Article 10
In-cabin voice systems and ADAS perception both classify as high-risk AI. Article 10 audits come down to three themes: data lineage, representativeness, and dataset-level bias examination. Vendors that produce evidence only on request cost you the audit window. YPAI ships the evidence pack at delivery, not on subpoena.
Perception edge-case retrain
A cyclist behind a parked van, a sensor calibration that disagrees by one frame between LiDAR cuboid and 2D box. The fix is not a bigger dataset; it is ASIL-aware curation of the long tail before it ships, with cross-sensor synchronisation that holds at 40 TB per hour of capture.
Five stages. Click any stage to see the actual artifact YPAI delivers at that point in the engagement.
Native-speaker capture in real cabin noise. Files delivered with every engagement:
audio_raw_48khz.wav speakers.json (demographic metadata) recording_environment.json (mic, room, ambient) consent_log.csv (per-speaker, GDPR Article 7) Each engagement specifies a coverage matrix. Example axes:
Idle 60 km/h 110 km/h Wipers on HVAC max Window down Passenger overlap Music ducked Sample count and gap analysis per cell delivered with the corpus.
Wake-word ground truth, intent slot, entity extraction, DMS state tag.
schema_version utterance_id audio_path wake_word_present (boolean + timestamp) intent (slot taxonomy) entities[] (per-utterance extraction) dms_state (drowsy / distracted / nominal) speaker_demographics Sample disagreement-resolution log fields:
batch_id sample_count inter_annotator_agreement disagreement_count reviewer_credentials resolution_log_path Audit-trail per batch. Available on procurement request.
Delivered with every engagement (not on subpoena). Article 10 high-risk AI documentation.
data_provenance.pdf bias_assessment.pdf consent_audit.csv risk_classification.md residency_attestation.pdf The free pilot covers recording AND annotation: two target languages, five hours of native-speaker in-cabin voice per language in real cabin noise, one thousand utterances per language with transcript and wake-word and intent labels. Delivered in five to seven business days. From there, the production engagement scopes by language coverage, scale, and regulatory context.
Native speakers, real cabin noise, 5-7 days delivery
Article 28 clauses pre-cleared, EEA-resident processing
Data lineage, consent framework, bias mitigation documented
After you submit the brief above, here is the timeline. The free pilot at T+5-7 days delivers real in-cabin voice recording against your target language and noise profile, not a deck.
T+1 day
A named EU-resident project lead replies within one business day with feasibility, language coverage, and a first read on Article 10 risk classification.
T+3 days
Indicative scope returned with target languages, cabin-noise capture profile (engine, HVAC, road, passenger overlap), wake-word spec, and Article 10 evidence pack manifest.
T+5-7 days
Free pilot covers recording AND annotation: 2 languages, 5h native-speaker recording per language in real cabin noise, 1000 utterances per language with transcript and wake-word and intent labels. Production engagement scopes from there by language coverage, scale, and regulatory context.
T+14 days
Article 28 clauses pre-cleared, EEA-resident processing committed in contract. Sub-processor list named, withdrawal SLA confirmed. Production recording begins.
Norwegian Aksjeselskap. EEA-resident operations. GDPR Article 7 consent on every contributor. EU AI Act Article 10 evidence pack at delivery.
Voice and perception data both classify as high-risk training material under EU AI Act Article 10. Both need documented lineage, consent provenance, and EEA-resident processing.
data_provenance.pdf bias_assessment.pdf bias_assessment.pdf consent_audit.csv consent_audit.csv dpa.pdf asil_taxonomy.pdf sotif_taxonomy.pdf cybersec_handling.pdf qa_process_log.pdf residency_attestation.pdf entity_attestation.pdf 30-day erasure SLA applies to both speaker recordings and perception assets. Sub-processor list and evidence artifacts delivered at engagement scoping.
The full automotive data surface in one engagement: in-cabin voice corpora, wake-word and DMS audio, perception annotation, real-world capture. EEA jurisdiction across all of it.
In-cabin voice corpora across 150+ languages (50+ productionised), wake-word and intent labeling, DMS audio cues, and TTS for branded voice personalities. Native-speaker capture in real cabin noise.
Cross-modal data collection and annotation pipelines that underpin both voice and perception programs: real-world capture, video and image annotation, Article 10 provenance throughout.
LiDAR, sensor fusion, and semantic segmentation when the voice engagement extends into the perception stack. ASIL-aware taxonomies, SOTIF edge-case coverage.
Telematics annotation and dashcam video work for fleet AI and predictive maintenance programs.
Add YPAI to your home screen
Tap the Share button, then Add to Home Screen.