The state of end-to-end systems for Mexican Spanish speech recognition

Carlos Daniel Hernández-Mena, Ivan Vladimir Meza Ruiz


Current end-to-end speech recognizer systems report an excellent performance for Spanish. However, this is not reported for specific variants. Moreover, it is unclear if there would be a benefit in creating a fine-tuned version for a particular variant. To investigate these aspects, particularly for Mexican Spanish, we evaluate four different of-the-shelf speech recognizers (one commercial and three open-source); additionally, we fine-tune two systems for Mexican Spanish. We evaluate read and spontaneous speech, present an error analysis and show that fine-tuning for a variant decreases the error rate. As a result of our experimentation, we build two new systems available to the community.

Texto completo: