We present an approach for unsupervised learning of speech representation disentangling contents and styles. Our model consists of: (1) a local encoder that captures per-frame information; (2) a global encoder that captures per-utterance information; and (3) a conditional decoder that reconstructs speech given local and global latent variables. Our experiments show that (1) the local latent variables encode speech contents, as reconstructed speech can be recognized by ASR with low word error rates (WER), even with a different global encoding; (2) the global latent variables encode speaker style, as reconstructed speech shares speaker identity with the source utterance of the global encoding. Additionally, we demonstrate an useful application from our pre-trained model, where we can train a speaker recognition model from the global latent variables and achieve high accuracy by fine-tuning with as few data as one label per speaker.
In this paper, we report our submitted system for the ZeroSpeech 2020 challenge on Track 2019. The main theme in this challenge is to build a speech synthesizer without any textual information or phonetic labels. In order to tackle those challenges, we build a system that must address two major components such as 1) given speech audio, extract subword units in an unsupervised way and 2) re-synthesize the audio from novel speakers. The system also needs to balance the codebook performance between the ABX error rate and the bitrate compression rate. Our main contribution here is we proposed Transformer-based VQ-VAE for unsupervised unit discovery and Transformer-based inverter for the speech synthesis given the extracted codebook. Additionally, we also explored several regularization methods to improve performance even further.
Deep acoustic models typically receive features in the first layer of the network, and process increasingly abstract representations in the subsequent layers. Here, we propose to feed the input features at multiple depths in the acoustic model. As our motivation is to allow acoustic models to re-examine their input features in light of partial hypotheses we introduce intermediate model heads and loss function. We study this architecture in the context of deep Transformer networks, and we use an attention mechanism over both the previous layer activations and the input features. To train this model's intermediate output hypothesis, we apply the objective function at each layer right before feature re-use. We find that the use of such intermediate losses significantly improves performance by itself, as well as enabling input feature re-use. We present results on both Librispeech, and a large scale video dataset, with relative improvements of 10 - 20% for Librispeech and 3.2 - 13% for videos.
We propose and evaluate transformer-based acoustic models (AMs) for hybrid speech recognition. Several modeling choices are discussed in this work, including various positional embedding methods and an iterated loss to enable training deep transformers. We also present a preliminary study of using limited right context in transformer models, which makes it possible for streaming applications. We demonstrate that on the widely used Librispeech benchmark, our transformer-based AM outperforms the best published hybrid result by 19% to 26% relative when the standard n-gram language model (LM) is used. Combined with neural network LM for rescoring, our proposed approach achieves state-of-the-art results on Librispeech. Our findings are also confirmed on a much larger internal dataset.
In this paper, we explore a method for training speech-to-speech translation tasks without any transcription or linguistic supervision. Our proposed method consists of two steps: First, we train and generate discrete representation with unsupervised term discovery with a discrete quantized autoencoder. Second, we train a sequence-to-sequence model that directly maps the source language speech to the target language's discrete representation. Our proposed method can directly generate target speech without any auxiliary or pre-training steps with a source or target transcription. To the best of our knowledge, this is the first work that performed pure speech-to-speech translation between untranscribed unknown languages.
The most common way for humans to communicate is by speech. But perhaps a language system cannot know what it is communicating without a connection to the real world by image perception. In fact, humans perceive these multiple sources of information together to build a general concept. However, constructing a machine that can alleviate these modalities together in a supervised learning fashion is difficult, because a parallel dataset is required among speech, image, and text modalities altogether that is often unavailable. A machine speech chain based on sequence-to-sequence deep learning was previously proposed to achieve semi-supervised learning that enabled automatic speech recognition (ASR) and text-to-speech synthesis (TTS) to teach each other when they receive unpaired data. In this research, we take a further step by expanding the speech chain into a multimodal chain and design a closely knit chain architecture that connects ASR, TTS, image captioning (IC), and image retrieval (IR) models into a single framework. ASR, TTS, IC, and IR components can be trained in a semi-supervised fashion by assisting each other given incomplete datasets and leveraging cross-modal data augmentation within the chain.
We describe our submitted system for the ZeroSpeech Challenge 2019. The current challenge theme addresses the difficulty of constructing a speech synthesizer without any text or phonetic labels and requires a system that can (1) discover subword units in an unsupervised way, and (2) synthesize the speech with a target speaker's voice. Moreover, the system should also balance the discrimination score ABX, the bit-rate compression rate, and the naturalness and the intelligibility of the constructed voice. To tackle these problems and achieve the best trade-off, we utilize a vector quantized variational autoencoder (VQ-VAE) and a multi-scale codebook-to-spectrogram (Code2Spec) inverter trained by mean square error and adversarial loss. The VQ-VAE extracts the speech to a latent space, forces itself to map it into the nearest codebook and produces compressed representation. Next, the inverter generates a magnitude spectrogram to the target voice, given the codebook vectors from VQ-VAE. In our experiments, we also investigated several other clustering algorithms, including K-Means and GMM, and compared them with the VQ-VAE result on ABX scores and bit rates. Our proposed approach significantly improved the intelligibility (in CER), the MOS, and discrimination ABX scores compared to the official ZeroSpeech 2019 baseline or even the topline.
The speech chain mechanism integrates automatic speech recognition (ASR) and text-to-speech synthesis (TTS) modules into a single cycle during training. In our previous work, we applied a speech chain mechanism as a semi-supervised learning. It provides the ability for ASR and TTS to assist each other when they receive unpaired data and let them infer the missing pair and optimize the model with reconstruction loss. If we only have speech without transcription, ASR generates the most likely transcription from the speech data, and then TTS uses the generated transcription to reconstruct the original speech features. However, in previous papers, we just limited our back-propagation to the closest module, which is the TTS part. One reason is that back-propagating the error through the ASR is challenging due to the output of the ASR are discrete tokens, creating non-differentiability between the TTS and ASR. In this paper, we address this problem and describe how to thoroughly train a speech chain end-to-end for reconstruction loss using a straight-through estimator (ST). Experimental results revealed that, with sampling from ST-Gumbel-Softmax, we were able to update ASR parameters and improve the ASR performances by 11\% relative CER reduction compared to the baseline.