Is ASR the right tool for the construction of Spoken Corpus Linguistics in European Spanish?

Resumen

Spoken corpora are a valuable resource to explore naturally occurring discourse. However, large parts of those corpora remain untranscribed due to the high cost of manually transcribing audio files; and, therefore, the access to these resources is limited. This problem could be faced by means of Automatic Speech Recognition (ASR) tools, that have shown their potential to automatically transcribe audio files. In this work, we study two families of ASR models (Whisper and Seamless) for automatically transcribing files from the COSER corpus (that stands for Corpus Oral y Sonoro del Español Rural, in English Audible Corpus of Rural Spanish). Our results show that those ASR models can produce accurate transcriptions independently of the dialect of the speakers and their speed-rate; specially with the large v3 version of Whisper that is the model which produces the best results (mean WER of 0.292). However, in some cases the transcriptions do not perfectly align with those produced by humans, since human transcriptors reflect nuances introduced in the speech of speakers that are not captured with the ASR models. This shows that ASR tools can reduce the burden of manually transcribing hours of audios from spoken corpus, but human supervision is still needed.