Synthetic speech is a useful source for augmentation of automatic speech recognition (ASR) systems, but there is a "sim-to-real" gap between synthetic and real speech that can limit generalization. The natural variability of real speech is essential to the training of robust ASR systems. While synthetic data augmentation can be used to approximate the variability of natural speech, however, not all aspects of variation are equally relevant for augmentation. In this work, we introduce speech re-painting, a method for in-context augmented synthesis, using target training datasets to generate new utterances guided by speech and text on the fly in a zero-shot manner. We evaluate this technique using downstream ASR word error rate (WER) using the VCTK and LibriSpeech datasets. These represent unique speaker and lexical challenges that are addressed by re-painting, realizing a reduction of WER more than 50% in particular settings.

Collecting high-quality studio recordings of audio is challenging, which limits the language coverage of text-to-speech (TTS) systems. This paper proposes a framework for scaling a multilingual TTS model to 100+ languages using found data without supervision. The proposed framework combines speech-text encoder pretraining with unsupervised training using untranscribed speech and unspoken text data sources, thereby leveraging massively multilingual joint speech and text representation learning. Without any transcribed speech in a new language, this TTS model can generate intelligible speech in >30 unseen languages (CER difference of <10% to ground truth). With just 15 minutes of transcribed, found data, we can reduce the intelligibility difference to 1% or less from the ground-truth, and achieve naturalness scores that match the ground-truth in several languages.

Although neural end-to-end text-to-speech models can synthesize highly natural speech, there is still room for improvements to its efficiency and naturalness. This paper proposes a non-autoregressive neural text-to-speech model augmented with a variational autoencoder-based residual encoder. This model, called Parallel Tacotron, is highly parallelizable during both training and inference, allowing efficient synthesis on modern parallel hardware. The use of the variational autoencoder relaxes the one-to-many mapping nature of the text-to-speech problem and improves naturalness. To further improve the naturalness, we use lightweight convolutions, which can efficiently capture local contexts, and introduce an iterative spectrogram loss inspired by iterative refinement. Experimental results show that Parallel Tacotron matches a strong autoregressive baseline in subjective evaluations with significantly decreased inference time.

This paper introduces Parallel Tacotron 2, a non-autoregressive neural text-to-speech model with a fully differentiable duration model which does not require supervised duration signals.The duration model is based on a novel attention mechanism and an iterative reconstruction loss based on Soft Dynamic Time Warping, this model can learn token-frame alignments as well as token durations automatically.Experimental results show that Parallel Tacotron 2 outperforms baselines in subjective naturalness in several diverse multi speaker evaluations.

BACKGROUND: Distance methods are ubiquitous tools in phylogenetics. Their primary purpose may be to reconstruct evolutionary history, but they are also used as components in bioinformatic pipelines. However, poor computational efficiency has been a constraint on the applicability of distance methods on very large problem instances. RESULTS: We present fastphylo, a software package containing implementations of efficient algorithms for two common problems in phylogenetics: estimating DNA/protein sequence distances and reconstructing a phylogeny from a distance matrix. We compare fastphylo with other neighbor joining based methods and report the results in terms of speed and memory efficiency. CONCLUSIONS: Fastphylo is a fast, memory efficient, and easy to use software suite. Due to its modular architecture, fastphylo is a flexible tool for many phylogenetic studies.

MOTIVATION: The evolution of viruses is very rapid and in addition to local point mutations (insertion, deletion, substitution) it also includes frequent recombinations, genome rearrangements and horizontal transfer of genetic materials (HGTS). Evolutionary analysis of viral sequences is therefore a complicated matter for two main reasons: First, due to HGTs and recombinations, the right model of evolution is a network and not a tree. Second, due to genome rearrangements, an alignment of the input sequences is not guaranteed. These facts encourage developing methods for inferring phylogenetic networks that do not require aligned sequences as input. RESULTS: In this work, we present the first computational approach which deals with both genome rearrangements and horizontal gene transfers and does not require a multiple alignment as input. We formalize a new set of computational problems which involve analyzing such complex models of evolution. We investigate their computational complexity, and devise algorithms for solving them. Moreover, we demonstrate the viability of our methods on several synthetic datasets as well as four biological datasets. AVAILABILITY: The code is available from the authors upon request.

BACKGROUND: Some distance methods are among the most commonly used methods for reconstructing phylogenetic trees from sequence data. The input to a distance method is a distance matrix, containing estimated pairwise distances between all pairs of taxa. Distance methods themselves are often fast, e.g., the famous and popular Neighbor Joining (NJ) algorithm reconstructs a phylogeny of n taxa in time O(n3). Unfortunately, the fastest practical algorithms known for Computing the distance matrix, from n sequences of length l, takes time proportional to l.n2. Since the sequence length typically is much larger than the number of taxa, the distance estimation is the bottleneck in phylogeny reconstruction. This bottleneck is especially apparent in reconstruction of large phylogenies or in applications where many trees have to be reconstructed, e.g., bootstrapping and genome wide applications. RESULTS: We give an advanced algorithm for Computing the number of mutational events between DNA sequences which is significantly faster than both Phylip and Paup. Moreover, we give a new method for estimating pairwise distances between sequences which contain ambiguity Symbols. This new method is shown to be more accurate as well as faster than earlier methods. CONCLUSION: Our novel algorithm for Computing distance estimators provides a valuable tool in phylogeny reconstruction. Since the running time of our distance estimation algorithm is comparable to that of most distance methods, the previous bottleneck is removed. All distance methods, such as NJ, require a distance matrix as input and, hence, our novel algorithm significantly improves the overall running time of all distance methods. In particular, we show for real world biological applications how the running time of phylogeny reconstruction using NJ is improved from a matter of hours to a matter of seconds.

Sorting permutations by transpositions is an important problem in genome rearrangements. A transposition is a rearrangement operation in which a segment is cut out of the permutation and pasted in a different location. The complexity of this problem is still open and it has been a 10-year-old open problem to improve the best known 1.5-approximation algorithm. In this paper, we provide a 1.375-approximation algorithm for sorting by transpositions. The algorithm is based on a new upper bound on the diameter of 3-permutations. In addition, we present some new results regarding the transposition diameter: we improve the lower bound for the transposition diameter of the symmetric group and determine the exact transposition diameter of simple permutations.