Accentus

Accentus is an American-English accent detection model launched around 2026-11-05 by Luel Lab, the research arm of voice-data company Luel. Model page: https://luel.work/lab/models/accentus ; benchmark: https://luel.work/lab/benchmarks/us-accent .

Purpose: when collecting voice data at scale, automatically verify whether contributors really have an American accent instead of manual listening.

Highlights:

• Built for real conversation and phone audio, not studio reads;
• Three outcomes: Pass (sounds American), Review (human listens — noise, phone lines, performed accents), Reject (not American, naming the group such as South Asian, West African, East African, Filipino);
• Under 10 seconds of detected speech: no score;
• Self-reported: only 5 of 233 Americans rejected (2.1%), ~5% of others let through; beats AccentCL, Qwen3.8 Omni Flash, Gemini 3.8 Flash on its US Accent board;
• Currently in closed beta.

Problem

Voice-data buyers often pay per accent/region; mislabeled clips poison datasets. Accentus uses bands, not percentages, sending uncertain cases to humans to avoid false rejects.

Design trade-offs

• Prefer Review over Reject: a wrong auto-reject costs more than human review;
• Explicitly says it cannot tell where a person is from — only accent features.

Limits & concerns

• Benchmark built and reported by Luel;
• Accent classification carries fairness/discrimination risk; use should be limited to data QA, not hiring or identity decisions;
• Closed beta, no public API yet.

Who it's for

• Voice-data collection/labeling platforms and ASR/TTS data teams;
• Voice AI companies filtering corpora by accent.

Getting started

1. Read https://luel.work/lab/models/accentus and the US Accent benchmark;
2. Request closed-beta access from Luel;
3. Feed at least 10 s (ideally 30 s) of detected speech and plan human review for the Review band;
4. Mind compliance: use for data quality control only.