In a range of recent works, object-centric architectures have been shown to be suitable for unsupervised scene decomposition in the vision domain. Inspired by these methods we present AudioSlots, a slot-centric generative model for blind source separation in the audio domain. AudioSlots is built using permutation-equivariant encoder and decoder networks. The encoder network based on the Transformer architecture learns to map a mixed audio spectrogram to an unordered set of independent source embeddings. The spatial broadcast decoder network learns to generate the source spectrograms from the source embeddings. We train the model in an end-to-end manner using a permutation invariant loss function. Our results on Libri2Mix speech separation constitute a proof of concept that this approach shows promise. We discuss the results and limitations of our approach in detail, and further outline potential ways to overcome the limitations and directions for future work.
We propose to characterize and improve the performance of blind room impulse response (RIR) estimation systems in the context of a downstream application scenario, far-field automatic speech recognition (ASR). We first draw the connection between improved RIR estimation and improved ASR performance, as a means of evaluating neural RIR estimators. We then propose a GAN-based architecture that encodes RIR features from reverberant speech and constructs an RIR from the encoded features, and uses a novel energy decay relief loss to optimize for capturing energy-based properties of the input reverberant speech. We show that our model outperforms the state-of-the-art baselines on acoustic benchmarks (by 72% on the energy decay relief and 22% on an early-reflection energy metric), as well as in an ASR evaluation task (by 6.9% in word error rate).
Speech emotion recognition (SER) classifies audio into emotion categories such as Happy, Angry, Fear, Disgust and Neutral. While Speech Emotion Recognition (SER) is a common application for popular languages, it continues to be a problem for low-resourced languages, i.e., languages with no pretrained speech-to-text recognition models. This paper firstly proposes a language-specific model that extract emotional information from multiple pre-trained speech models, and then designs a multi-domain model that simultaneously performs SER for various languages. Our multidomain model employs a multi-gating mechanism to generate unique weighted feature combination for each language, and also searches for specific neural network structure for each language through a neural architecture search module. In addition, we introduce a contrastive auxiliary loss to build more separable representations for audio data. Our experiments show that our model raises the state-of-the-art accuracy by 3% for German and 14.3% for French.
Filler words like ``um" or ``uh" are common in spontaneous speech. It is desirable to automatically detect and remove them in recordings, as they affect the fluency, confidence, and professionalism of speech. Previous studies and our preliminary experiments reveal that the biggest challenge in filler word detection is that fillers can be easily confused with other hard categories like ``a" or ``I". In this paper, we propose a novel filler word detection method that effectively addresses this challenge by adding auxiliary categories dynamically and applying an additional inter-category focal loss. The auxiliary categories force the model to explicitly model the confusing words by mining hard categories. In addition, inter-category focal loss adaptively adjusts the penalty weight between ``filler" and ``non-filler" categories to deal with other confusing words left in the ``non-filler" category. Our system achieves the best results, with a huge improvement compared to other methods on the PodcastFillers dataset.
Self-supervised learning (SSL) speech models have achieved unprecedented success in speech representation learning, but some questions regarding their representation ability remain unanswered. This paper addresses two of them: (1) Can SSL speech models deal with non-speech audio?; (2) Would different SSL speech models have insights into diverse aspects of audio features? To answer the two questions, we conduct extensive experiments on abundant speech and non-speech audio datasets to evaluate the representation ability of currently state-of-the-art SSL speech models, which are wav2vec 2.0 and HuBERT in this paper. These experiments are carried out during NeurIPS 2021 HEAR Challenge as a standard evaluation pipeline provided by competition officials. Results show that (1) SSL speech models could extract meaningful features of a wide range of non-speech audio, while they may also fail on certain types of datasets; (2) different SSL speech models have insights into different aspects of audio features. The two conclusions provide a foundation for the ensemble of representation models. We further propose an ensemble framework to fuse speech representation models' embeddings. Our framework outperforms state-of-the-art SSL speech/audio models and has generally superior performance on abundant datasets compared with other teams in HEAR Challenge. Our code is available at https://github.com/tony10101105/HEAR-2021-NeurIPS-Challenge -- NTU-GURA.
Speech recognition applications cover a range of different audio and text distributions, with different speaking styles, background noise, transcription punctuation and character casing. However, many speech recognition systems require dataset-specific tuning (audio filtering, punctuation removal and normalisation of casing), therefore assuming a-priori knowledge of both the audio and text distributions. This tuning requirement can lead to systems failing to generalise to other datasets and domains. To promote the development of multi-domain speech systems, we introduce the End-to-end Speech Benchmark (ESB) for evaluating the performance of a single automatic speech recognition (ASR) system across a broad set of speech datasets. Benchmarked systems must use the same data pre- and post-processing algorithm across datasets - assuming the audio and text data distributions are a-priori unknown. We compare a series of state-of-the-art (SoTA) end-to-end (E2E) systems on this benchmark, demonstrating how a single speech system can be applied and evaluated on a wide range of data distributions. We find E2E systems to be effective across datasets: in a fair comparison, E2E systems achieve within 2.6% of SoTA systems tuned to a specific dataset. Our analysis reveals that transcription artefacts, such as punctuation and casing, pose difficulties for ASR systems and should be included in evaluation. We believe E2E benchmarking over a range of datasets promotes the research of multi-domain speech recognition systems. ESB is available at https://huggingface.co/esb.
How to boost speech pre-training with textual data is an unsolved problem due to the fact that speech and text are very different modalities with distinct characteristics. In this paper, we propose a cross-modal Speech and Language Model (SpeechLM) to explicitly align speech and text pre-training with a pre-defined unified discrete representation. Specifically, we introduce two alternative discrete tokenizers to bridge the speech and text modalities, including phoneme-unit and hidden-unit tokenizers, which can be trained using a small amount of paired speech-text data. Based on the trained tokenizers, we convert the unlabeled speech and text data into tokens of phoneme units or hidden units. The pre-training objective is designed to unify the speech and the text into the same discrete semantic space with a unified Transformer network. Leveraging only 10K text sentences, our SpeechLM gets a 16\% relative WER reduction over the best base model performance (from 6.8 to 5.7) on the public LibriSpeech ASR benchmark. Moreover, SpeechLM with fewer parameters even outperforms previous SOTA models on CoVoST-2 speech translation tasks. We also evaluate our SpeechLM on various spoken language processing tasks under the universal representation evaluation framework SUPERB, demonstrating significant improvements on content-related tasks. Our code and models are available at https://aka.ms/SpeechLM.
Speech emotion recognition (SER) is the task of recognising human's emotional states from speech. SER is extremely prevalent in helping dialogue systems to truly understand our emotions and become a trustworthy human conversational partner. Due to the lengthy nature of speech, SER also suffers from the lack of abundant labelled data for powerful models like deep neural networks. Pre-trained complex models on large-scale speech datasets have been successfully applied to SER via transfer learning. However, fine-tuning complex models still requires large memory space and results in low inference efficiency. In this paper, we argue achieving a fast yet effective SER is possible with self-distillation, a method of simultaneously fine-tuning a pretrained model and training shallower versions of itself. The benefits of our self-distillation framework are threefold: (1) the adoption of self-distillation method upon the acoustic modality breaks through the limited ground-truth of speech data, and outperforms the existing models' performance on an SER dataset; (2) executing powerful models at different depth can achieve adaptive accuracy-efficiency trade-offs on resource-limited edge devices; (3) a new fine-tuning process rather than training from scratch for self-distillation leads to faster learning time and the state-of-the-art accuracy on data with small quantities of label information.
Recently, the ever-increasing demand for bandwidth in multi-modal communication systems requires a paradigm shift. Powered by deep learning, semantic communications are applied to multi-modal scenarios to boost communication efficiency and save communication resources. However, the existing end-to-end neural network (NN) based framework without the channel encoder/decoder is incompatible with modern digital communication systems. Moreover, most end-to-end designs are task-specific and require re-design and re-training for new tasks, which limits their applications. In this paper, we propose a distributed multi-modal semantic communication framework incorporating the conventional channel encoder/decoder. We adopt NN-based semantic encoder and decoder to extract correlated semantic information contained in different modalities, including speech, text, and image. Based on the proposed framework, we further establish a general rate-adaptive coding mechanism for various types of multi-modal semantic tasks. In particular, we utilize unequal error protection based on semantic importance, which is derived by evaluating the distortion bound of each modality. We further formulate and solve an optimization problem that aims at minimizing inference delay while maintaining inference accuracy for semantic tasks. Numerical results show that the proposed mechanism fares better than both conventional communication and existing semantic communication systems in terms of task performance, inference delay, and deployment complexity.
Text-to-speech synthesis (TTS) is a task to convert texts into speech. Two of the factors that have been driving TTS are the advancements of probabilistic models and latent representation learning. We propose a TTS method based on latent variable conversion using a diffusion probabilistic model and the variational autoencoder (VAE). In our TTS method, we use a waveform model based on VAE, a diffusion model that predicts the distribution of latent variables in the waveform model from texts, and an alignment model that learns alignments between the text and speech latent sequences. Our method integrates diffusion with VAE by modeling both mean and variance parameters with diffusion, where the target distribution is determined by approximation from VAE. This latent variable conversion framework potentially enables us to flexibly incorporate various latent feature extractors. Our experiments show that our method is robust to linguistic labels with poor orthography and alignment errors.