Speech and text are two major forms of human language. The research community has been focusing on mapping speech to text or vice versa for many years. However, in the field of language modeling, very little effort has been made to model them jointly. In light of this, we explore joint language modeling for speech units and text. Specifically, we compare different speech tokenizers to transform continuous speech signals into discrete units and use different methods to construct mixed speech-text data. We introduce automatic metrics to evaluate how well the joint LM mixes speech and text. We also fine-tune the LM on downstream spoken language understanding (SLU) tasks with different modalities (speech or text) and test its performance to assess the model's learning of shared representations. Our results show that by mixing speech units and text with our proposed mixing techniques, the joint LM improves over a speech-only baseline on SLU tasks and shows zero-shot cross-modal transferability.
Recent work on speech representation models jointly pre-trained with text has demonstrated the potential of improving speech representations by encoding speech and text in a shared space. In this paper, we leverage such shared representations to address the persistent challenge of limited data availability in spoken language understanding tasks. By employing a pre-trained speech-text model, we find that models fine-tuned on text can be effectively transferred to speech testing data. With as little as 1 hour of labeled speech data, our proposed approach achieves comparable performance on spoken language understanding tasks (specifically, sentiment analysis and named entity recognition) when compared to previous methods using speech-only pre-trained models fine-tuned on 10 times more data. Beyond the proof-of-concept study, we also analyze the latent representations. We find that the bottom layers of speech-text models are largely task-agnostic and align speech and text representations into a shared space, while the top layers are more task-specific.
Speech enhancement systems are typically trained using pairs of clean and noisy speech. In audio-visual speech enhancement (AVSE), there is not as much ground-truth clean data available; most audio-visual datasets are collected in real-world environments with background noise and reverberation, hampering the development of AVSE. In this work, we introduce AV2Wav, a resynthesis-based audio-visual speech enhancement approach that can generate clean speech despite the challenges of real-world training data. We obtain a subset of nearly clean speech from an audio-visual corpus using a neural quality estimator, and then train a diffusion model on this subset to generate waveforms conditioned on continuous speech representations from AV-HuBERT with noise-robust training. We use continuous rather than discrete representations to retain prosody and speaker information. With this vocoding task alone, the model can perform speech enhancement better than a masking-based baseline. We further fine-tune the diffusion model on clean/noisy utterance pairs to improve the performance. Our approach outperforms a masking-based baseline in terms of both automatic metrics and a human listening test and is close in quality to the target speech in the listening test. Audio samples can be found at https://home.ttic.edu/~jcchou/demo/avse/avse_demo.html.
Many self-supervised speech models (S3Ms) have been introduced over the last few years, producing performance and data efficiency improvements for a variety of speech tasks. Evidence is emerging that different S3Ms encode linguistic information in different layers, and also that some S3Ms appear to learn phone-like sub-word units. However, the extent to which these models capture larger linguistic units, such as words, and where word-related information is encoded, remains unclear. In this study, we conduct several analyses of word segment representations extracted from different layers of three S3Ms: wav2vec2, HuBERT, and WavLM. We employ canonical correlation analysis (CCA), a lightweight analysis tool, to measure the similarity between these representations and word-level linguistic properties. We find that the maximal word-level linguistic content tends to be found in intermediate model layers, while some lower-level information like pronunciation is also retained in higher layers of HuBERT and WavLM. Syntactic and semantic word attributes have similar layer-wise behavior. We also find that, for all of the models tested, word identity information is concentrated near the center of each word segment. We then test the layer-wise performance of the same models, when used directly with no additional learned parameters, on several tasks: acoustic word discrimination, word segmentation, and semantic sentence similarity. We find similar layer-wise trends in performance, and furthermore, find that when using the best-performing layer of HuBERT or WavLM, it is possible to achieve performance on word segmentation and sentence similarity that rivals more complex existing approaches.
State-of-the-art text-to-speech (TTS) systems require several hours of recorded speech data to generate high-quality synthetic speech. When using reduced amounts of training data, standard TTS models suffer from speech quality and intelligibility degradations, making training low-resource TTS systems problematic. In this paper, we propose a novel extremely low-resource TTS method called Voice Filter that uses as little as one minute of speech from a target speaker. It uses voice conversion (VC) as a post-processing module appended to a pre-existing high-quality TTS system and marks a conceptual shift in the existing TTS paradigm, framing the few-shot TTS problem as a VC task. Furthermore, we propose to use a duration-controllable TTS system to create a parallel speech corpus to facilitate the VC task. Results show that the Voice Filter outperforms state-of-the-art few-shot speech synthesis techniques in terms of objective and subjective metrics on one minute of speech on a diverse set of voices, while being competitive against a TTS model built on 30 times more data.
Any-to-any voice conversion (VC) aims to convert the timbre of utterances from and to any speakers seen or unseen during training. Various any-to-any VC approaches have been proposed like AUTOVC, AdaINVC, and FragmentVC. AUTOVC, and AdaINVC utilize source and target encoders to disentangle the content and speaker information of the features. FragmentVC utilizes two encoders to encode source and target information and adopts cross attention to align the source and target features with similar phonetic content. Moreover, pre-trained features are adopted. AUTOVC used dvector to extract speaker information, and self-supervised learning (SSL) features like wav2vec 2.0 is used in FragmentVC to extract the phonetic content information. Different from previous works, we proposed S2VC that utilizes Self-Supervised features as both source and target features for VC model. Supervised phoneme posteriororgram (PPG), which is believed to be speaker-independent and widely used in VC to extract content information, is chosen as a strong baseline for SSL features. The objective evaluation and subjective evaluation both show models taking SSL feature CPC as both source and target features outperforms that taking PPG as source feature, suggesting that SSL features have great potential in improving VC.
The few-shot multi-speaker multi-style voice cloning task is to synthesize utterances with voice and speaking style similar to a reference speaker given only a few reference samples. In this work, we investigate different speaker representations and proposed to integrate pretrained and learnable speaker representations. Among different types of embeddings, the embedding pretrained by voice conversion achieves the best performance. The FastSpeech 2 model combined with both pretrained and learnable speaker representations shows great generalization ability on few-shot speakers and achieved 2nd place in the one-shot track of the ICASSP 2021 M2VoC challenge.
Prosody modeling is an essential component in modern text-to-speech (TTS) frameworks. By explicitly providing prosody features to the TTS model, the style of synthesized utterances can thus be controlled. However, predicting natural and reasonable prosody at inference time is challenging. In this work, we analyzed the behavior of non-autoregressive TTS models under different prosody-modeling settings and proposed a hierarchical architecture, in which the prediction of phoneme-level prosody features are conditioned on the word-level prosody features. The proposed method outperforms other competitors in terms of audio quality and prosody naturalness in our objective and subjective evaluation.