A script for the Soundflow platform which automates the process of loudness analysis in the Dolby Atmos Album Assembler. I wrote this function alongside Dolby Atmos mix engineer Luke Argilla to address a pain point in his workflow. The impetus for this project was that for projects with either a large number of tracks or particularly long tracks, the process of individually analyzing each track can be prohibitively time consuming. With this script, an Atmos mastering engineer can set up their DA3 session and walk away to do other work while every track in the session is analyzed. The code for this project is hosted on my Github. Feel free to use it, tweak it, and let me know if it's helped you at all!

Word error rate (WER) is a commonly used metric for determining the efficacy of automatic speech recognition models. The simple formula for word error rate tallies substitutions, insertions, and deletions and divides the total number of errors by the number of words in the reference transcription (S+D+I / N). This is computationally simple, but it fails to capture effective quality of the predicted transcript because it treats all errors as equal.

This python notebook demonstrates an experimental "Information Weighted Word Error Rate" which weights traditional WER with semantic relevance of mispredicted words. After computing a word position aligned matrix of all substitutions, insertions, and deletions using the Levenshtein Distance Algorithm, the Gemini API is used to generate a weight vector for the semantic relevance of each word. This weight vector is multiplied with the SDI matrix and the IWWER is calculated as the sum of all nonzero elements / number of words in the reference transcript.

The code for this project is hosted on my Github. It's more of an exploration and not a suitable for practical use in training a model, but feel free to check it out!