Existing resources for Automatic Speech Recognition in Portuguese are mostly focused on Brazilian Portuguese, leaving European Portuguese (EP) and other varieties underexplored. To bridge this gap, we introduce CAMÕES, the first open framework for EP and other Portuguese varieties. It consists of (1) a comprehensive evaluation benchmark, including 46 h of EP test data spanning multiple domains; and (2) a collection of state-of-the-art models. For the latter, we consider multiple foundation models, evaluating their zero-shot and fine-tuned performances, as well as E-Branchformer models trained from scratch. A curated set of $\mathbf{4 2 5 h}$ of EP was used for both fine-tuning and training. Our results show comparable performance for EP between fine-tuned foundation models and the E-Branchformer. Furthermore, the best-performing models achieve relative improvements above 35% WER, compared to the strongest zero-shot foundation model, establishing a new state-of-the-art for EP and other varieties.

The Multi-Level Alignment System (MuLAS) is the L2F tool for building multi-tier speech corpora with reduced or no human intervention at all. MuLAS automatically combines information coming from external speech annotations, human or machine-generated, with the text-based utterance descriptions that it creates, in order to build more reliable and complete descriptions of the spoken utterances. This paper presents our methods for multi-tier annotation synchronization, which lie behind the MuLAS operation. Such methods have allowed us to expand the building of multi-tier corpora to new languages without spending too much effort. MuLAS has been successfully applied to the building of multitier corpora for speech synthesis in American and British English, European Portuguese and German. Natural prosody generation has benefited from MuLAS, too, since prosodic models can be derived from corpora built by MuLAS.

This paper describes a speech segmentation tool allowing alternative word pronunciations within a WFST framework. Two approaches to word pronunciation graph generation were developed and evaluated. The first approach is grapheme-based where each grapheme is converted into all the phones it can give rise to, in the form of a WFST. Word graphs are obtained by concatenating all grapheme WFSTs. In the second approach, a training corpus is used to find the different realizations of the syllable. This information is used to generate alternative syllable-level pronunciations, represented as WFSTs, that are concatenated to produce the word graphs. Both approaches were evaluated by aligning the phone sequence generated by each approach with the manually labelled phone sequence for all utterance in the corpus. This alignment was used for computing F-rate values for each phone. The syllable-based approach produced the best results.

In this paper we describe a system designed to be used on a shop floor of a factory that produces non-flat ceramic products, where an extensive mixture of human and automatic labour is present. Today manufacturing setups rely increasingly on technology. It is common to have all sources of equipment on the shop floor, commanded by industrial PCs or PLCs connected by an industrial network to other factory resources. Also, the production systems are becoming more and more autonomous requiring less operator intervention in every day normal operation. That means using computers for controlling and supervision of the production systems, industrial networks and distributed software architectures. It means also designing application software that is really distributed in the shop floor, taking advantage of the flexibility installed by using programmable equipment.

This paper presents the results of our effort in improving the accuracy of a DTW-based automatic phonetic aligner. The adopted model assumes that the phonetic segment sequence is already known and so the goal is only to align the spoken utterance with a reference synthetic signal produced by waveform concatenation without prosodic modifications. Instead of using a single acoustic measure to compute the alignment cost function, our strategy uses a combination of acoustic features depending on the pair of phonetic segment classes being aligned. The results show that this strategy considerably reduces the segment boundary location errors, even when aligning synthetic and natural speech signals of different gender speakers.