Speech is a natural and preferred means of expressing one’s thoughts and emotions for communication purpose. Speech and language impairments are negatively impacting the daily life of a large population worldwide. Speech impairments are manifested in the aspects of atypical articulation and phonation, while language impairments could be present across multiple linguistic levels in the use of spoken or written language. Timely and reliable assessment on the type and severity of impairment is crucial to effective treatment and rehabilitation. Conventionally speech assessment is carried out by professional speech and language pathologists (SLPs). In view of the shortage of qualified SLPs with relevant linguistic and cultural background, objective assessment techniques based on acoustical signal analysis and machine learning models are expected to play an increasingly important role in assisting clinical assessment. This presentation will cover a series of our recent studies on applying deep learning models to automatic assessment of different types of speech and language impairments. The types of impairments that we have tackled include voice disorder in adults, phonology and articulation disorder in children, and neurological disorder in elderly people. All of our works are focused on spoken Cantonese. The use of Siamese network and auto-encoder model has been investigated to address the challenges related to the scarcity of training speech and the absence of reliable labels. The findings in attempting the end-to-end approach to speech assessment will also be shared.

Even conversational systems have attracted a lot of attention recently, the current systems sometimes fail due to the errors from different components. This talk presents potential directions for improvement: 1) we first focus on learning language embeddings specifically for practical scenarios for better robustness, and 2) secondly we propose a novel learning framework for natural language understanding and generation on top of duality for better scalability. Both directions enhance the robustness and scalability of conversational systems, showing the potential of guiding future research areas.

Real-world environments are always contain stationary and/or time-varying noises that are received together with speech signals by recording devices. The received noises inevitably degrade the performance of human--human and human--machine interfaces, and this issue has attracted significant attention over the years. To address this issue, an important front-end speech process, namely speech enhancement, which extracts clean components from noisy input, can improve the voice quality and intelligibility of noise-deteriorated clean speech. These speech-enhancement systems can be split into two categories in terms of the physical configurations: single- and multi-channels. For single-channel-based speech enhancement systems, the speech waveform was recorded essentially from an microphone, and then enhanced through the enhancement system, which is derived based on the temporal information of the input. Multiple microphones are used to record the input speech in a multi-channel-based speech enhancement system. The system is designed by simultaneously exploiting the spatial diversity and temporal structures of received signals. In this talk, we present our recent research achievements using machine learning and signal processing on improving speech perception abilities for both configurations.

Speech signal carries a lot of redundant information for speech understanding, and many studies have showed that the loss of some acoustic information did not significantly affect speech intelligibility if important acoustic information was preserved. Due to their hearing loss, hearing-impaired listeners are unable to recognize some acoustic information (e.g., temporal fine structure). Hence, studying the important acoustic information minimally required for an intelligible speech in different listening environments could guide our design of novel assistive hearing technologies. In this talk, I will first introduce early work on the relative importance of commonly-used acoustic cues for speech intelligibility, particularly on a vocoder model for speech intelligibility. Then, I will present recent studies towards reconstructing an intelligible speech with cortical EEG signals, including Mandarin tone imagery and speech reconstruction.

The fields of clinical otology and hearing research are advancing at the forefront of innovation in medicine and technology. Promising progress in genetic medicine and digital technology have started to change the traditional medical and hearing research. Next-generation sequencing, novel gene therapy vectors, CRISPR-Cas9 gene editing technologies, mobile-phone apps and artificial intelligence all generates enormous creative energy. In this talk, I will introduce how these revolutionary technologies change physician’s practice and research.

As privacy is getting more and more concerned, voice user interface (UI) is in the process of transition from the cloud to the edge device. However, to land a neural network based voice/language model on edge devices with efficient power consumption is very challenging. In this talk, we will first introduce MediaTek NeuroPilot from the platform level to tackle this challenge. Then, more specifically, we investigate it from the algorithm perspective including the algorithm trend and landing opportunity. Finally, we show some preliminary results on speech recognitions and natural language understanding.

Voice disorders mainly result from chronic overuse or abuse, particularly for teachers or other occupational voice users. Previous studies have proposed a contact microphone attached to the anterior neck for ambulatory voice monitoring; however, the inconvenience associated with taping and wiring, and the lack of real-time processing has limited its daily application.

Starting from 2015, we founded a research group collaborating with experts from National Yang-Ming University, Yuan Ze University and Far Eastern Memorial Hospital. We proposed an system using wireless microphone for real-time ambulatory voice monitoring. We invited 10 teachers to participate in the pilot study. We designed an adaptive threshold (AT) function to detect the presence of speech based on energy envelope. All the participant wore a wireless microphone during a teaching class (around 40-60 minutes), in quite classroom (background noise < 55dB SPL). We developed a software for manually labeling speech segments according to the time and frequency domains. We randomly selected 25 utterance (10 s each) from the recorded audio files for calculating the coefficients for AT function via genetic algorithm. Another five random utterances were used for testing the accuracy of ASD system, using manually labeled data as the ground truth. We measured phonation ratio (speech frames / total frames) and the length of speech segments as a proxy of phonation habits of the users. We also mimicked scenarios of noisy backgrounds by manually mixing 4 different types of noise into the original recordings. Adjuvant noise reduction function using Log MMSE algorithm was applied to counteract the influence of detection accuracy.

The study results exhibited detection accuracy (for speech) ranging from 81% to 94%. Subsequent analyses revealed a phonation ratio between 50% and 78%, with most phonation segments less than 10 s. Although the presence of background noise reduced the accuracy of the ASD system (25% to 79%), adjuvant noise reduction function can effectively improve the accuracy for up to 45.8%, especially under stable noise (e.g. white noise).

study demonstrated a good detection accuracy of the proposed system. Preliminary results of phonation ratio and speech segments were all comparable to those of previous research. Although wireless microphone was susceptible to background noise, additional noise reduction function can overcome this limitation. These results indicate that the proposed system can be applied to ambulatory voice monitoring for occupational voice users.