Large Linguistic Corpus Reduction with SCP Algorithms

Abstract

Linguistic corpus design is a critical concern for building rich annotated corpora useful in different domains of applications. For example, speech technologies such as ASR (Automatic Speech Recognition) or TTS (Text-to-Speech) need a huge amount of speech data to train data-driven models or to produce synthetic speech. Collecting data is always related to costs (recording speech, verifying annotations, etc.), and as a rule of thumb, the more data you gather, the more costly your application will be. Within this context, we present in this article solutions to reduce the amount of linguistic text content while maintaining a sufficient level of linguistic richness required by a model or an application. This problem can be formalized as a Set Covering Problem (SCP) and we evaluate two algorithmic heuristics applied to design large text corpora in English and French for covering phonological information or POS labels. The first considered algorithm is a standard greedy solution with an agglomerative/spitting strategy and we propose a second algorithm based on Lagrangian relaxation. The latter approach provides a lower bound to the cost of each covering solution. This lower bound can be used as a metric to evaluate the quality of a reduced corpus whatever the algorithm applied. Experiments show that a suboptimal algorithm like a greedy algorithm achieves good results; the cost of its solutions is not so far from the lower bound (about 4.35% for 3-phoneme coverings). Usually, constraints in SCP are binary; we proposed here a generalization where the constraints on each covering feature can be multi-valued.

Corpus Design Using Convolutional Auto-Encoder Embeddings for Audio-Book Synthesis

“…• SC: this system is based on a set covering problem which is solved by a greedy strategy [5]. The best utterances are selected to cover η times each linguistic feature.…”

“…The most commonly used algorithmic strategy is the greedy one. In [5], the combination of agglomerative and spitting greedy phases have been assessed and compared to a Lagrangian relaxation based algorithm to derive full multirepresented coverings of diphoneme and triphonemes. The Lagrangian relaxation approach provides a lower bound showing that greedy strategies build solutions close to optimal ones.…”

“…The second one is based on a greedy strategy to design an utterance subset close to a target linguistic distribution. These selection approaches are compared to a standard set covering one as the covering of all diphonemes using a greedy strategy [5,3]. Objective and subjective results show that these proposed methods are more efficient than the standard one.…”

Corpus design for expressive speech: impact of the utterance length

“…The main proposed approaches consist in extracting, from a large textual corpus, for instance the target book to be vocalized, a minimal subset of sentences that maximizes an optimisation criterion. This criterion is often related to the maximization of the linguistic coverage [4,5,6] (formalized as a set covering problem) or the closeness to a target linguistic distribution [7,8]. Different algorithms have been compared and the mainly used approach is the greedy one, providing a good trade-off between the computational time and closeness to the optimal solution [6].…”

“…Some studies, as in [6], point out that the designed voices tend to be composed of utterances shorter than those of the initial pool. In [4], the set covering problem is dealt with a greedy strategy which selects a sub-corpus with an average length of 20 phonemes per sentence out of an initial corpus with an average length of 74 (this approach will be named set covering in the following).…”