Auditoria: new open-source speech model prioritizes dialect diversity for inclusive ASR

AI · 5 min read

Auditoria, an open-source automatic speech recognition model, was unveiled today with an explicit mandate: reduce historic bias in transcription accuracy for nonstandard dialects, regional accents, and disfluent speech. The model was trained on a mix of controlled corpora and community-contributed audio, with participants retaining granular consent controls over how clips are used and shared.

The team published a new evaluation suite that measures performance across over 40 dialect groups and includes metrics for error types that disproportionately affect intelligibility for assistive use cases (e.g., proper nouns, code-switching, and filled pauses). Early benchmarks show substantial error-rate reductions for Caribbean English and South Asian English varieties compared with mainstream commercial models.

Privacy and data governance were central to the project: contributors can opt into training-only uses, and the release includes reproducible training recipes so organizations can fine-tune models on local, consented datasets. Accessibility advocates see Auditoria as a promising step toward more inclusive assistive tools, but stressed ongoing funding and community maintenance will be essential to keep datasets representative.