We experiment with a low-latency, end-to-end voice-to-voice communication model to optimize it for real-time conversational applications. By analyzing components essential to voice to voice (V-2-V) system viz. automatic speech recognition (ASR), text-to-speech (TTS), and dialog management, our work analyzes how to reduce processing time while maintaining high-quality interactions to identify the levers for optimizing V-2-V system. Our work identifies that TTS component which generates life-like voice, full of emotions including natural pauses and exclamations has highest impact on Real time factor (RTF). The experimented V-2-V architecture utilizes CSM1b has the capability to understand tone as well as context of conversation by ingesting both audio and text of prior exchanges to generate contextually accurate speech. We explored optimization of Residual Vector Quantization (RVQ) iterations by the TTS decoder which come at a cost of decrease in the quality of voice generated. Our experimental evaluations also demonstrate that for V-2-V implementations based on CSM most important optimizations can be brought by reducing the number of RVQ Iterations along with the codebooks used in Mimi.

Transformer-based neural speech processing has achieved state-of-the-art performance. Since speech audio signals are known to be highly compressible, here we seek to accelerate neural speech transcription by time-domain signal sparsification early in the neural encoding stage, taking advantage of the interpretability of the self-attention mechanism in transformer audio encoders. With the Whisper family of models, we perform a systematic architecture search over the joint space of sparsification stage (a certain encoder layer) and compression ratio (sparsity). We found that the best resulting solutions under 1% accuracy degradation choose to sparsify the hidden state to 40-60% sparsity at an early encoding stage, and thereby achieve up to 1.6x runtime acceleration in English speech transcription tasks on Nvidia GPUs without any fine-tuning.

Predictive maintenance has been used to optimize system repairs in the industrial, medical, and financial domains. This technique relies on the consistent ability to detect and predict anomalies in critical systems. AI models have been trained to detect system faults, improving predictive maintenance efficiency. Typically there is a lack of fault data to train these models, due to organizations working to keep fault occurrences and down time to a minimum. For newly installed systems, no fault data exists since they have yet to fail. By using diffusion models for synthetic data generation, the complex training datasets for these predictive models can be supplemented with high level synthetic fault data to improve their performance in anomaly detection. By learning the relationship between healthy and faulty data in similar systems, a diffusion model can attempt to apply that relationship to healthy data of a newly installed system that has no fault data. The diffusion model would then be able to generate useful fault data for the new system, and enable predictive models to be trained for predictive maintenance. The following paper demonstrates a system for generating useful, multivariate synthetic data for predictive maintenance, and how it can be applied to systems that have yet to fail.

In multi-view medical diagnosis, deep learning-based models often fuse information from different imaging perspectives to improve diagnostic performance. However, existing approaches are prone to overfitting and rely heavily on view-specific features, which can lead to trivial solutions. In this work, we introduce Random Token Fusion (RTF), a novel technique designed to enhance multi-view medical image analysis using vision transformers. By integrating randomness into the feature fusion process during training, RTF addresses the issue of overfitting and enhances the robustness and accuracy of diagnostic models without incurring any additional cost at inference. We validate our approach on standard mammography and chest X-ray benchmark datasets. Through extensive experiments, we demonstrate that RTF consistently improves the performance of existing fusion methods, paving the way for a new generation of multi-view medical foundation models.

We approach designing a state-space model for deep learning applications through its dual representation, the transfer function, and uncover a highly efficient sequence parallel inference algorithm that is state-free: unlike other proposed algorithms, state-free inference does not incur any significant memory or computational cost with an increase in state size. We achieve this using properties of the proposed frequency domain transfer function parametrization, which enables direct computation of its corresponding convolutional kernel's spectrum via a single Fast Fourier Transform. Our experimental results across multiple sequence lengths and state sizes illustrates, on average, a 35% training speed improvement over S4 layers -- parametrized in time-domain -- on the Long Range Arena benchmark, while delivering state-of-the-art downstream performances over other attention-free approaches. Moreover, we report improved perplexity in language modeling over a long convolutional Hyena baseline, by simply introducing our transfer function parametrization. Our code is available at https://github.com/ruke1ire/RTF.

Scale has opened new frontiers in natural language processing, but at a high cost. In response, by learning to only activate a subset of parameters in training and inference, Mixture-of-Experts (MoE) have been proposed as an energy efficient path to even larger and more capable language models and this shift towards a new generation of foundation models is gaining momentum, particularly within the field of Automatic Speech Recognition (ASR). Recent works that incorporating MoE into ASR models have complex designs such as routing frames via supplementary embedding network, improving multilingual ability for the experts, and utilizing dedicated auxiliary losses for either expert load balancing or specific language handling. We found that delicate designs are not necessary, while an embarrassingly simple substitution of MoE layers for all Feed-Forward Network (FFN) layers is competent for the ASR task. To be more specific, we benchmark our proposed model on a large scale inner-source dataset (160k hours), the results show that we can scale our baseline Conformer (Dense-225M) to its MoE counterparts (MoE-1B) and achieve Dense-1B level Word Error Rate (WER) while maintaining a Dense-225M level Real Time Factor (RTF). Furthermore, by applying Unified 2-pass framework with bidirectional attention decoders (U2++), we achieve the streaming and non-streaming decoding modes in a single MoE based model, which we call U2++ MoE. We hope that our study can facilitate the research on scaling speech foundation models without sacrificing deployment efficiency.

Recent progress in large-scale zero-shot speech synthesis has been significantly advanced by language models and diffusion models. However, the generation process of both methods is slow and computationally intensive. Efficient speech synthesis using a lower computing budget to achieve quality on par with previous work remains a significant challenge. In this paper, we present FlashSpeech, a large-scale zero-shot speech synthesis system with approximately 5\% of the inference time compared with previous work. FlashSpeech is built on the latent consistency model and applies a novel adversarial consistency training approach that can train from scratch without the need for a pre-trained diffusion model as the teacher. Furthermore, a new prosody generator module enhances the diversity of prosody, making the rhythm of the speech sound more natural. The generation processes of FlashSpeech can be achieved efficiently with one or two sampling steps while maintaining high audio quality and high similarity to the audio prompt for zero-shot speech generation. Our experimental results demonstrate the superior performance of FlashSpeech. Notably, FlashSpeech can be about 20 times faster than other zero-shot speech synthesis systems while maintaining comparable performance in terms of voice quality and similarity. Furthermore, FlashSpeech demonstrates its versatility by efficiently performing tasks like voice conversion, speech editing, and diverse speech sampling. Audio samples can be found in https://flashspeech.github.io/.

This paper reports our work on building up a Cantonese Speech-to-Text (STT) system with a syllable based acoustic model. This is a part of an effort in building a STT system to aid dyslexic students who have cognitive deficiency in writing skills but have no problem expressing their ideas through speech. For Cantonese speech recognition, the basic unit of acoustic models can either be the conventional Initial-Final (IF) syllables, or the Onset-Nucleus-Coda (ONC) syllables where finals are further split into nucleus and coda to reflect the intra-syllable variations in Cantonese. By using the Kaldi toolkit, our system is trained using the stochastic gradient descent optimization model with the aid of GPUs for the hybrid Deep Neural Network and Hidden Markov Model (DNN-HMM) with and without I-vector based speaker adaptive training technique. The input features of the same Gaussian Mixture Model with speaker adaptive training (GMM-SAT) to DNN are used in all cases. Experiments show that the ONC-based syllable acoustic modeling with I-vector based DNN-HMM achieves the best performance with the word error rate (WER) of 9.66% and the real time factor (RTF) of 1.38812.

Automatic Speech Recognition (ASR) systems are used in the financial domain to enhance the caller experience by enabling natural language understanding and facilitating efficient and intuitive interactions. Increasing use of ASR systems requires that such systems exhibit very low error rates. The predominant ASR models to collect numeric data are large, general-purpose commercial models -- Google Speech-to-text (STT), or Amazon Transcribe -- or open source (OpenAI's Whisper). Such ASR models are trained on hundreds of thousands of hours of audio data and require considerable resources to run. Despite recent progress large speech recognition models, we highlight the potential of smaller, specialized "micro" models. Such light models can be trained perform well on number recognition specific tasks, competing with general models like Whisper or Google STT while using less than 80 minutes of training time and occupying at least an order of less memory resources. Also, unlike larger speech recognition models, micro-models are trained on carefully selected and curated datasets, which makes them highly accurate, agile, and easy to retrain, while using low compute resources. We present our work on creating micro models for multi-digit number recognition that handle diverse speaking styles reflecting real-world pronunciation patterns. Our work contributes to domain-specific ASR models, improving digit recognition accuracy, and privacy of data. An added advantage, their low resource consumption allows them to be hosted on-premise, keeping private data local instead uploading to an external cloud. Our results indicate that our micro-model makes less errors than the best-of-breed commercial or open-source ASRs in recognizing digits (1.8% error rate of our best micro-model versus 5.8% error rate of Whisper), and has a low memory footprint (0.66 GB VRAM for our model versus 11 GB VRAM for Whisper).

Classical methods for acoustic scene mapping require the estimation of time difference of arrival (TDOA) between microphones. Unfortunately, TDOA estimation is very sensitive to reverberation and additive noise. We introduce an unsupervised data-driven approach that exploits the natural structure of the data. Our method builds upon local conformal autoencoders (LOCA) - an offline deep learning scheme for learning standardized data coordinates from measurements. Our experimental setup includes a microphone array that measures the transmitted sound source at multiple locations across the acoustic enclosure. We demonstrate that LOCA learns a representation that is isometric to the spatial locations of the microphones. The performance of our method is evaluated using a series of realistic simulations and compared with other dimensionality-reduction schemes. We further assess the influence of reverberation on the results of LOCA and show that it demonstrates considerable robustness.