Abstract
Speech translation has been traditionally tackled under a cascade approach, chaining speech recognition and machine translation components to translate from an audio source in a given language into text or speech in a target language. Leveraging on deep learning approaches to natural language processing, recent studies have explored the potential of direct end-to-end neural modelling to perform the speech translation task. Though several benefits may come from end-to-end modelling, such as a reduction in latency and error propagation, the comparative merits of each approach still deserve detailed evaluations and analyses. In this work, we compared state-of-the-art cascade and direct approaches on the under-resourced Basque–Spanish language pair, which features challenging phenomena such as marked differences in morphology and word order. This case study thus complements other studies in the field, which mostly revolve around the English language. We describe and analysed in detail the mintzai-ST corpus, prepared from the sessions of the Basque Parliament, and evaluated the strengths and limitations of cascade and direct speech translation models trained on this corpus, with variants exploiting additional data as well. Our results indicated that, despite significant progress with end-to-end models, which may outperform alternatives in some cases in terms of automated metrics, a cascade approach proved optimal overall in our experiments and manual evaluations.
Highlights
Introduction published maps and institutional affilSpeech translation (ST) systems have been traditionally designed under a cascade approach, where independent automatic speech recognition (ASR) and machine translation (MT) components are chained, feeding the ASR output into the MT component, oftentimes with task-specific bridging to optimise component communication [1,2,3]
The remainder of this paper is organised as follows: Section 2 presents related work in the field; in Section 3, we describe the mintzai-ST corpus, including the data acquisition process and data statistics; Section 4 describes the different baseline models that were built for Basque–Spanish speech translation, including cascade and end-to-end models; Section 5 discusses comparative results for the baseline models; in Section 6, we describe several direct ST model variants and their results on automated metrics; Section 7 describes the protocol and results of our manual evaluation of the best cascade and end-to-end models, along with the results of targeted evaluations on specific linguistic phenomena and on the impact of relative input difficulty; Section 8 draws the main conclusions from this work
ASR models trained with either an End-to-End neural model (E2E) or the Kaldi toolkit (KAL); ASR and MT models trained on either In-Domain data only (IND) or on a combination of in-domain and out-of-domain data (ALL), by integrating the OpenDataEuskadi dataset to train the language and casing models for speech recognition and the translation models for the MT component; MT models obtained by fine-tuning a model trained on the out-of-domain dataset with the in-domain data, in addition to the models trained on in-domain data only and all available data
Summary
Standard speech-to-text translation systems operate on the basis of separate components for speech recognition and machine translation, feeding the output of the ASR module into the MT component. One of the main reasons for this state of affairs was training data scarcity, i.e., the lack of sufficiently large speech-to-text datasets to train direct ST systems, in contrast with the comparatively larger training data for the ASR and MT components, considered separately Another relevant factor was the need to improve end-to-end ST architectures or training methods for this type of approach. Recent improvements in ST modelling have closed the gap between direct and cascade approaches on standard datasets Whereas the latter outperformed the former in the IWSLT 2019 shared task, results from the 2020 edition featured similar performances overall [23].
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
