Automatic Dialectal Transcription: An Evaluation on Finnish and Norwegian

The fields of dialectology and sociolinguistics are highly reliant on phonetically transcribed spoken language data, which are often written in language-specific styles and alphabets. Transcribing dialectal speech phonetically is time-consuming and thus a major bottleneck for data collection in variational linguistics. Meanwhile, the field of automatic speech recognition (ASR) has taken leaps forward in recent years. In this work, we introduce automatic dialectal transcription as a distinct ASR task and investigate solutions using data from two unrelated languages (Finnish and Norwegian) with two levels of transcription precision. We find a large performance gap between the languages, as Finnish models are much more efficient than those for Norwegian. We further evaluate the character-level errors of both languages and dialectal difficulties of the best Finnish model. We find that the most geographically central dialects tend to be easier to transcribe than the more distant ones.