Cascade or Direct Speech Translation? A Case Study

Speech translation has been traditionally tackled under a cascade approach, chaining speech recognition and machine components to translate from an audio source in given language into text or target language. Leveraging on deep learning approaches natural processing, recent studies have explored the potential of direct end-to-end neural modelling perform task. Though several benefits may come modelling, such as reduction latency error propagation, comparative merits each approach still deserve detailed evaluations analyses. In this work, we compared state-of-the-art under-resourced Basque–Spanish pair, which features challenging phenomena marked differences morphology word order. This case study thus complements other field, mostly revolve around English We describe analysed detail mintzai-ST corpus, prepared sessions Basque Parliament, evaluated strengths limitations models trained with variants exploiting additional data well. Our results indicated that, despite significant progress models, outperform alternatives some cases terms automated metrics, proved optimal overall our experiments manual evaluations.