Self-supervised learning leverages unlabeled data effectively, improving label efficiency and generalization to domains without labeled data. While recent work has studied generalization to more acoustic/linguistic domains, languages, and modalities, these investigations are limited to single-source speech with one primary speaker in the recording. This paper presents Cocktail HuBERT, a self-supervised learning framework that generalizes to mixture speech using a masked pseudo source separation objective. This objective encourages the model to identify the number of sources, separate and understand the context, and infer the content of masked regions represented as discovered units. Cocktail HuBERT outperforms state-of-the-art results with 69% lower WER on multi-speaker ASR, 31% lower DER on diarization, and is competitive on single- and multi-speaker tasks from SUPERB.
The rapid advancement of deep neural networks has significantly improved various tasks, such as image and speech recognition. However, as the complexity of these models increases, so does the computational cost and the number of parameters, making it difficult to deploy them on resource-constrained devices. This paper proposes a novel memorization-based inference (MBI) that is compute free and only requires lookups. Specifically, our work capitalizes on the inference mechanism of the recurrent attention model (RAM), where only a small window of input domain (glimpse) is processed in a one time step, and the outputs from multiple glimpses are combined through a hidden vector to determine the overall classification output of the problem. By leveraging the low-dimensionality of glimpse, our inference procedure stores key value pairs comprising of glimpse location, patch vector, etc. in a table. The computations are obviated during inference by utilizing the table to read out key-value pairs and performing compute-free inference by memorization. By exploiting Bayesian optimization and clustering, the necessary lookups are reduced, and accuracy is improved. We also present in-memory computing circuits to quickly look up the matching key vector to an input query. Compared to competitive compute-in-memory (CIM) approaches, MBI improves energy efficiency by almost 2.7 times than multilayer perceptions (MLP)-CIM and by almost 83 times than ResNet20-CIM for MNIST character recognition.
Target speaker information can be utilized in speech enhancement (SE) models to more effectively extract the desired speech. Previous works introduce the speaker embedding into speech enhancement models by means of concatenation or affine transformation. In this paper, we propose a speaker attentive module to calculate the attention scores between the speaker embedding and the intermediate features, which are used to rescale the features. By merging this module in the state-of-the-art SE model, we construct the personalized SE model for ICASSP Signal Processing Grand Challenge: DNS Challenge 5 (2023). Our system achieves a final score of 0.529 on the blind test set of track1 and 0.549 on track2.
In this work, we propose a frequency bin-wise method to estimate the single-channel speech presence probability (SPP) with multiple deep neural networks (DNNs) in the short-time Fourier transform domain. Since all frequency bins are typically considered simultaneously as input features for conventional DNN-based SPP estimators, high model complexity is inevitable. To reduce the model complexity and the requirements on the training data, we take a single frequency bin and some of its neighboring frequency bins into account to train separate gate recurrent units. In addition, the noisy speech and the a posteriori probability SPP representation are used to train our model. The experiments were performed on the Deep Noise Suppression challenge dataset. The experimental results show that the speech detection accuracy can be improved when we employ the frequency bin-wise model. Finally, we also demonstrate that our proposed method outperforms most of the state-of-the-art SPP estimation methods in terms of speech detection accuracy and model complexity.
End-to-End speech-to-speech translation (S2ST) is generally evaluated with text-based metrics. This means that generated speech has to be automatically transcribed, making the evaluation dependent on the availability and quality of automatic speech recognition (ASR) systems. In this paper, we propose a text-free evaluation metric for end-to-end S2ST, named BLASER, to avoid the dependency on ASR systems. BLASER leverages a multilingual multimodal encoder to directly encode the speech segments for source input, translation output and reference into a shared embedding space and computes a score of the translation quality that can be used as a proxy to human evaluation. To evaluate our approach, we construct training and evaluation sets from more than 40k human annotations covering seven language directions. The best results of BLASER are achieved by training with supervision from human rating scores. We show that when evaluated at the sentence level, BLASER correlates significantly better with human judgment compared to ASR-dependent metrics including ASR-SENTBLEU in all translation directions and ASR-COMET in five of them. Our analysis shows combining speech and text as inputs to BLASER does not increase the correlation with human scores, but best correlations are achieved when using speech, which motivates the goal of our research. Moreover, we show that using ASR for references is detrimental for text-based metrics.
LLMs have demonstrated remarkable abilities at interacting with humans through language, especially with the usage of instruction-following data. Recent advancements in LLMs, such as MiniGPT-4, LLaVA, and X-LLM, further enlarge their abilities by incorporating multi-modal inputs, including image, video, and speech. Despite their effectiveness at generating precise and detailed language understanding of the given modality signal, these LLMs give up the ability to ground specific parts of inputs, thus only constructing a coarse-grained mapping. However, explicit and informative correspondence between text and other modalities will not only improve the user experience but also help to expand the application scenario of multi-modal LLMs. Therefore, we propose BuboGPT, a multi-modal LLM with visual grounding that can perform cross-modal interaction between vision, audio and language, providing fine-grained understanding of visual objects and other given modalities. As a result, BuboGPT is able to point out the specific location of an object in the image, when it is generating response or description for that object. Our contributions are two-fold: 1) An off-the-shelf visual grounding module based on SAM that extracts entities in a sentence and find corresponding masks in the image. 2) A two-stage training scheme and instruction dataset to endow joint text-image-audio understanding. Our experiments show that BuboGPT achieves impressive multi-modality understanding and visual grounding abilities during the interaction with human. It performs consistently well when provided by arbitrary modality combinations (either aligned or unaligned). Our code, model and dataset are available at https://bubo-gpt.github.io .
Speech-to-speech translation directly translates a speech utterance to another between different languages, and has great potential in tasks such as simultaneous interpretation. State-of-art models usually contains an auxiliary module for phoneme sequences prediction, and this requires textual annotation of the training dataset. We propose a direct speech-to-speech translation model which can be trained without any textual annotation or content information. Instead of introducing an auxiliary phoneme prediction task in the model, we propose to use bottleneck features as intermediate training objectives for our model to ensure the translation performance of the system. Experiments on Mandarin-Cantonese speech translation demonstrate the feasibility of the proposed approach and the performance can match a cascaded system with respect of translation and synthesis qualities.
Silent speech interface is a promising technology that enables private communications in natural language. However, previous approaches only support a small and inflexible vocabulary, which leads to limited expressiveness. We leverage contrastive learning to learn efficient lipreading representations, enabling few-shot command customization with minimal user effort. Our model exhibits high robustness to different lighting, posture, and gesture conditions on an in-the-wild dataset. For 25-command classification, an F1-score of 0.8947 is achievable only using one shot, and its performance can be further boosted by adaptively learning from more data. This generalizability allowed us to develop a mobile silent speech interface empowered with on-device fine-tuning and visual keyword spotting. A user study demonstrated that with LipLearner, users could define their own commands with high reliability guaranteed by an online incremental learning scheme. Subjective feedback indicated that our system provides essential functionalities for customizable silent speech interactions with high usability and learnability.
To translate speech for automatic dubbing, machine translation needs to be isochronous, i.e. translated speech needs to be aligned with the source in terms of speech durations. We introduce target factors in a transformer model to predict durations jointly with target language phoneme sequences. We also introduce auxiliary counters to help the decoder to keep track of the timing information while generating target phonemes. We show that our model improves translation quality and isochrony compared to previous work where the translation model is instead trained to predict interleaved sequences of phonemes and durations.
Speech-driven gesture synthesis is a field of growing interest in virtual human creation. However, a critical challenge is the inherent intricate one-to-many mapping between speech and gestures. Previous studies have explored and achieved significant progress with generative models. Notwithstanding, most synthetic gestures are still vastly less natural. This paper presents DiffMotion, a novel speech-driven gesture synthesis architecture based on diffusion models. The model comprises an autoregressive temporal encoder and a denoising diffusion probability Module. The encoder extracts the temporal context of the speech input and historical gestures. The diffusion module learns a parameterized Markov chain to gradually convert a simple distribution into a complex distribution and generates the gestures according to the accompanied speech. Compared with baselines, objective and subjective evaluations confirm that our approach can produce natural and diverse gesticulation and demonstrate the benefits of diffusion-based models on speech-driven gesture synthesis.