Virtual Race Audio Commentary Generation

Generated commentary combined with externally created visuals

Replaced the previous outsourced commentary solution that had poor quality and required extensive QA.
Built the replacement from scratch in short timescales, resulting in commentary of reliably higher quality, increased flexibility, diagnostic reports and far less QA effort required.
Algorithmically generated race commentary audio by combining race metadata of runner positions with pre-recorded audio clips of runner numbers, positions and phrases.
Rejection of phrase choices with number position changes, preventing incorrect reads without failing previously valid results.
Automatic retry of races that failed all section retries, using alternative phrase combinations.
Ensured whole system remained deterministic despite the apparent random generation of the audio content.

Technologies: Java, Eclipse, macOS, ffmpeg

Provided increased usage of the recorded script phrases than the previous outsourced solution, improving the variation of commentary generated.
Reduced production turnaround of phrase and script changes from two weeks to several minutes.
After script and source audio proving runs, the QA rejection of race commentary issues was reduced from 30% to zero. This was for production runs of over two thousand race events with over thirty hours of video content.
Single solution worked for multiple race types.

Generated commentary combined with externally created visuals