2024 · Published journal article

Transformer-Based Turkish Automatic Speech Recognition

An evaluation of self-supervised speech representations and transfer learning for Turkish speech recognition.

Research question

What carries over when a speech model learns Turkish?

Evidence and scope

With Kutan Koruyan and Cihan Çılgın, I adapted a Wav2Vec 2.0 model pretrained on speech in 53 languages. We used recordings and transcripts from Mozilla Common Voice Turkish to teach the model to turn Turkish speech into text.

The paper reports a 23% word error rate on the Common Voice Turkish test split. This measure counts substitutions, deletions and insertions relative to the length of the reference transcript.

Primary source

Contribution
Davut Emre Taşar · First author
Source attribution
Davut Emre Taşar, Kutan Koruyan, Cihan Çılgın
Publication status
Published journal article
Publication or platform
Acta Infologica
Identifier
10.26650/acin.1338604

Sources

Research areas · Language

Related work