artificial-polyglot/arti

A Server for checking correctness of Bible audio files , especially for low resource languages # Artificial Polyglot (arti) An Application Server for checking correctness of Bible audio files, especially for low resource languages. ## NOTE: This repository is in active development. The production version of this application can be found at github.com ## Objective Artificial Polyglot is an open source software service that validates audio Bibles by using AI to compare audio recordings against the original text. Developed for Faith Comes by Hearing, it has reduced the time required to proof an audio New Testament from 10-15 days to 2-3 hours. The service, which is nicknamed Arti, was developed pro bono and is MIT-licensed. It is currently used to validate audio Bibles that are recorded in the field. ## Features * Read Text - Supports USX, USFM, Excel script, and CSV formats. All text absent from the audio recordings is automatically removed, such as footnote and cross references. * Read Audio - Supports MP3, and WAV files, organized either as chapter files, or script line files. * Verse Delineation - Aeneas (for comparison only), MMS Fairseq forced alignment, and MMS model forced alignment. * MMS Adapter Model Training - Trains an MMS adapter model from the provided text and audio recording. Samples that forced alignment detected that were likely errors are excluded from training. * MMS Adapter Inference - Uses the trained adapter model to generate a transcription for an audio recording. * MMS Model inference - Generates a transcription using the base MMS model without additional training, by searching a language graph to find the closest language that MMS supports. * Difference Report - Produces a character-level difference report with an integrated audio player, enabling human reviewers to identify and confirm errors. * Output Delivery - Results are stored, and delivered by email to the requesting user. ### Input Formats Supported **Text:** * USFM - Unified Standard Format, also called SFM * USX - Unified …