Text-to-speech (TTS) systems have achieved substantial improvements in quality with the use of neural architectures. Despite these quality improvements, they remain computationally intensive, which limits their deployment on low-power devices. This thesis explores whether a custom hardware accelerator can improve the efficiency of end-to-end TTS inference while maintaining or improving synthesis quality.

A phoneme-to-mel model based on EfficientSpeech is adapted with hardware-aware modifications, such as the replacement of expensive normalisation layers and simplification of activations. These modifications reduce the computational complexity while improving synthesis quality compared to the baseline model. The HiFi-GAN vocoder is optimised by reducing its model size from 920k to 296k parameters, while maintaining a fair quality. Combined, the complete TTS pipeline is reduced from 1.2M parameters to 562K, demonstrating that substantial reductions in both model size and complexity are achievable without sacrificing the synthesis quality.

A custom instruction set and hardware accelerator are designed to support the essential operations used in TTS inference, which include matrix operations, one-dimensional convolutions, and activation functions. The accelerator was synthesised and implemented on an FPGA using an out-of-context mode, achieving a post-implementation clock period of 4.182 ns, while reporting a power consumption of 1.569 W. The reported power consumption is approximately 9.9 times lower than the used baseline GPU system.

An end-to-end TTS inference accelerator is successfully implemented and verified, achieving a consistent real-time factor (RTF) of approximately 0.156 over varying input lengths, confirming real-time inference capability. For the longest test input of 31 phonemes, the accelerator uses 11.3 times less energy than the baseline system, despite the slightly higher RTF compared to the baseline’s 0.172. For shorter inputs, the energy efficiency improves further, with the accelerator consuming over 50 times less energy for a 4-phoneme test case.

This work demonstrates that dedicated hardware accelerators combined with hardware-aware model optimisations can significantly improve the efficiency of real-time neural TTS, enabling deployment on low-power devices such as embedded systems and battery-powered platforms that typically have strict power budgets and cannot rely on continuous mains power.

Audio samples of the optimised TTS models can be found at a demo page, and the hardware accelerator implementation is available open-source at the specified repository.

This paper presents a study focused on developing an efficient signal processing pipeline and identifying suitable machine learning models for real-time gesture recognition using a testbed consisting of an Arduino Nano 33 BLE and three OPT101 photodiodes. Our research aims to address the challenges of limited computational power whilst maintaining a high inference accuracy.

Experiments were conducted to optimise the signal processing and explore various machine learning model architectures, specifically revolving around convolutional neural networks. The data used for these experiments was gathered by creating a dataset of gestures from left- and right-handed participants. We took ethical considerations regarding participant recruitment and data security into account and we made sure to balance the dataset with both left- and right-handed participants as much as possible.

We obtained accurate gesture recognition results, surpassing the goal of a 75% success rate. Our machine learning models, trained on pre-processed 2D data, achieved near real-time inference times while running on the resource-constrained Arduino Nano 33 BLE.

The findings of this study contribute to the field of gesture recognition by providing insights into efficient signal processing techniques and identifying suitable machine learning models for resource-constrained devices. The developed system can be applied in various applications, ranging from games to healthcare. Furthermore, a dataset is contributed which can be used for further research.