Converting legacy media to text to increase findability of cultural records

Abstract

Many recordings made in the past have little metadata and so are difficult to find by normal catalog searches. We have been exploring the use of language models for automated speech recognition (ASR) with very small training sets, in one case (Bislama) 29 hours of transcribed audio, and, in the other (Nafsan) 10 hours. We will show that the error rate is remarkably low (5% for Bislama, 10% for Nafsan) and that, having developed the model, the ASR can run on a laptop computer with no internet connection. The tool, wri4en by Aso Mahmudi, called Easper, produces Elan files, segmented and transcribed. A second tool, wri4en by Raphaël Merx, takes a Bislama transcription in Elan format and translates it from Bislama to English. In each case some correction is still needed, but even the raw transcript provides a point of entry to the recording. The plan we are working on now with PARADISEC is to test language models against items in other languages in the collection in order to create a first-pass transcription. We will then have an ingestion pipeline for untranscribed media that automatically creates a textual transcript which will provide a searchable index of the media. We are building an online manual transcription system that will allow users to correct automatically generated transcripts.