Search

Recently, multi-stage systems have stood out among deep learning-based speech enhancement methods. However, these systems are always high in complexity, requiring millions of parameters and powerful computational resources, which limits their application for real-time processing in low-power devices. Besides, the contribution of various influencing factors to the success of multi-stage systems remains unclear, which presents challenges to reduce the size of these systems. In this paper, we extensively investigate a lightweight two-stage network with only 560k total parameters. It consists of a Mel-scale magnitude masking model in the first stage and a complex spectrum mapping model in the second stage. We first provide a consolidated view of the roles of gain power factor, post-filter, and training labels for the Mel-scale masking model. Then, we explore several training schemes for the two-stage network and provide some insights into the superiority of the two-stage network. We show that the proposed two-stage network trained by an optimal scheme achieves a performance similar to a four times larger open source model DeepFilterNet2., Comment: To appear at ICASSP 2024

Previous methods for predicting room acoustic parameters and speech quality metrics have focused on the single-channel case, where room acoustics and Mean Opinion Score (MOS) are predicted for a single recording device. However, quality-based device selection for rooms with multiple recording devices may benefit from a multi-channel approach where the descriptive metrics are predicted for multiple devices in parallel. Following our hypothesis that a model may benefit from multi-channel training, we develop a multi-channel model for joint MOS and room acoustics prediction (MOSRA) for five channels in parallel. The lack of multi-channel audio data with ground truth labels necessitated the creation of simulated data using an acoustic simulator with room acoustic labels extracted from the generated impulse responses and labels for MOS generated in a student-teacher setup using a wav2vec2-based MOS prediction model. Our experiments show that the multi-channel model improves the prediction of the direct-to-reverberation ratio, clarity, and speech transmission index over the single-channel model with roughly 5$\times$ less computation while suffering minimal losses in the performance of the other metrics., Comment: Accepted at ICASSP 2024

Deep learning models have become widely adopted in various domains, but their performance heavily relies on a vast amount of data. Datasets often contain a large number of irrelevant or redundant samples, which can lead to computational inefficiencies during the training. In this work, we introduce, for the first time in the context of the audio domain, the k-means clustering as a method for efficient data pruning. K-means clustering provides a way to group similar samples together, allowing the reduction of the size of the dataset while preserving its representative characteristics. As an example, we perform clustering analysis on the keyword spotting (KWS) dataset. We discuss how k-means clustering can significantly reduce the size of audio datasets while maintaining the classification performance across neural networks (NNs) with different architectures. We further comment on the role of scaling analysis in identifying the optimal pruning strategies for a large number of samples. Our studies serve as a proof-of-principle, demonstrating the potential of data selection with distance-based clustering algorithms for the audio domain and highlighting promising research avenues.

The recent ubiquitous adoption of remote conferencing has been accompanied by omnipresent frustration with distorted or otherwise unclear voice communication. Audio enhancement can compensate for low-quality input signals from, for example, small true wireless earbuds, by applying noise suppression techniques. Such processing relies on voice activity detection (VAD) with low latency and the added capability of discriminating the wearer's voice from others - a task of significant computational complexity. The tight energy budget of devices as small as modern earphones, however, requires any system attempting to tackle this problem to do so with minimal power and processing overhead, while not relying on speaker-specific voice samples and training due to usability concerns. This paper presents the design and implementation of a custom research platform for low-power wireless earbuds based on novel, commercial, MEMS bone-conduction microphones. Such microphones can record the wearer's speech with much greater isolation, enabling personalized voice activity detection and further audio enhancement applications. Furthermore, the paper accurately evaluates a proposed low-power personalized speech detection algorithm based on bone conduction data and a recurrent neural network running on the implemented research platform. This algorithm is compared to an approach based on traditional microphone input. The performance of the bone conduction system, achieving detection of speech within 12.8ms at an accuracy of 95\% is evaluated. Different SoC choices are contrasted, with the final implementation based on the cutting-edge Ambiq Apollo 4 Blue SoC achieving 2.64mW average power consumption at 14uJ per inference, reaching 43h of battery life on a miniature 32mAh li-ion cell and without duty cycling.

Since the mental states of the speaker modulate speech, stress introduced by cognitive or physical loads could be detected in the voice. The existing voice stress detection benchmark has shown that the audio embeddings extracted from the Hybrid BYOL-S self-supervised model perform well. However, the benchmark only evaluates performance separately on each dataset, but does not evaluate performance across the different types of stress and different languages. Moreover, previous studies found strong individual differences in stress susceptibility. This paper presents the design and development of voice stress detection, trained on more than 100 speakers from 9 language groups and five different types of stress. We address individual variabilities in voice stress analysis by adding speaker embeddings to the hybrid BYOL-S features. The proposed method significantly improves voice stress detection performance with an input audio length of only 3-5 seconds., Comment: 5 pages, 2 figures. Accepted at Interspeech 2023

This paper presents ALO-VC, a non-parallel low-latency one-shot phonetic posteriorgrams (PPGs) based voice conversion method. ALO-VC enables any-to-any voice conversion using only one utterance from the target speaker, with only 47.5 ms future look-ahead. The proposed hybrid signal processing and machine learning pipeline combines a pre-trained speaker encoder, a pitch predictor to predict the converted speech's prosody, and positional encoding to convey the phoneme's location information. We introduce two system versions: ALO-VC-R, which uses a pre-trained d-vector speaker encoder, and ALO-VC-E, which improves performance using the ECAPA-TDNN speaker encoder. The experimental results demonstrate both ALO-VC-R and ALO-VC-E can achieve comparable performance to non-causal baseline systems on the VCTK dataset and two out-of-domain datasets. Furthermore, both proposed systems can be deployed on a single CPU core with 55 ms latency and 0.78 real-time factor. Our demo is available online., Comment: Accepted to Interspeech 2023. Some audio samples are available at https://bohan7.github.io/ALO-VC-demo/

Estimating the quality of remote speech communication is a complex task influenced by the speaker, transmission channel, and listener. For example, the degradation of transmission quality can increase listeners' cognitive load, which can influence the overall perceived quality of the conversation. This paper presents a framework that isolates quality-dependent changes and controls most outside influencing factors like personal preference in a simulated conversational environment. The performed statistical analysis finds significant relationships between stimulus quality and the listener's valence and personality (agreeableness and openness) and, similarly, between the perceived task load during the listening task and the listener's personality and frustration intolerance. The machine learning model of the task load prediction improves the correlation coefficients from 0.48 to 0.76 when listeners' individuality is considered. The proposed evaluation framework and results pave the way for personalized audio quality assessment that includes speakers' and listeners' individuality beyond conventional channel modeling., Comment: IEEE ICASSP 2023 Conference

Voice Activity Detection (VAD) is a fundamental module in many audio applications. Recent state-of-the-art VAD systems are often based on neural networks, but they require a computational budget that usually exceeds the capabilities of a small battery-operated device when preserving the performance of larger models. In this work, we rely on the input from a bone conduction microphone (BCM) to design an efficient VAD (BC-VAD) robust against residual non-stationary noises originating from the environment or speakers not wearing the BCM.We first show that a larger VAD system (58k parameters) achieves state-of-the-art results on a publicly available benchmark but fails when running on bone conduction signals. We then compare its variant BC-VAD (5k parameters and trained on BC data) with a baseline especially designed for a BCM and show that the proposed method achieves better performances under various metrics while keeping the realtime processing requirement for a microcontroller.

Automatic speech quality assessment is essential for audio researchers, developers, speech and language pathologists, and system quality engineers. The current state-of-the-art systems are based on framewise speech features (hand-engineered or learnable) combined with time dependency modeling. This paper proposes an efficient system with results comparable to the best performing model in the ConferencingSpeech 2022 challenge. Our proposed system is characterized by a smaller number of parameters (40-60x), fewer FLOPS (100x), lower memory consumption (10-15x), and lower latency (30x). Speech quality practitioners can therefore iterate much faster, deploy the system on resource-limited hardware, and, overall, the proposed system contributes to sustainable machine learning. The paper also concludes that framewise embeddings outperform utterance-level embeddings and that multi-task training with acoustic conditions modeling does not degrade speech quality prediction while providing better interpretation.

Methods for extracting audio and speech features have been studied since pioneering work on spectrum analysis decades ago. Recent efforts are guided by the ambition to develop general-purpose audio representations. For example, deep neural networks can extract optimal embeddings if they are trained on large audio datasets. This work extends existing methods based on self-supervised learning by bootstrapping, proposes various encoder architectures, and explores the effects of using different pre-training datasets. Lastly, we present a novel training framework to come up with a hybrid audio representation, which combines handcrafted and data-driven learned audio features. All the proposed representations were evaluated within the HEAR NeurIPS 2021 challenge for auditory scene classification and timestamp detection tasks. Our results indicate that the hybrid model with a convolutional transformer as the encoder yields superior performance in most HEAR challenge tasks., Comment: Submitted to HEAR-PMLR 2021

The acoustic environment can degrade speech quality during communication (e.g., video call, remote presentation, outside voice recording), and its impact is often unknown. Objective metrics for speech quality have proven challenging to develop given the multi-dimensionality of factors that affect speech quality and the difficulty of collecting labeled data. Hypothesizing the impact of acoustics on speech quality, this paper presents MOSRA: a non-intrusive multi-dimensional speech quality metric that can predict room acoustics parameters (SNR, STI, T60, DRR, and C50) alongside the overall mean opinion score (MOS) for speech quality. By explicitly optimizing the model to learn these room acoustics parameters, we can extract more informative features and improve the generalization for the MOS task when the training data is limited. Furthermore, we also show that this joint training method enhances the blind estimation of room acoustics, improving the performance of current state-of-the-art models. An additional side-effect of this joint prediction is the improvement in the explainability of the predictions, which is a valuable feature for many applications., Comment: Submitted to Interspeech 2022

As a neurophysiological response to threat or adverse conditions, stress can affect cognition, emotion and behaviour with potentially detrimental effects on health in the case of sustained exposure. Since the affective content of speech is inherently modulated by an individual's physical and mental state, a substantial body of research has been devoted to the study of paralinguistic correlates of stress-inducing task load. Historically, voice stress analysis (VSA) has been conducted using conventional digital signal processing (DSP) techniques. Despite the development of modern methods based on deep neural networks (DNNs), accurately detecting stress in speech remains difficult due to the wide variety of stressors and considerable variability in the individual stress perception. To that end, we introduce a set of five datasets for task load detection in speech. The voice recordings were collected as either cognitive or physical stress was induced in the cohort of volunteers, with a cumulative number of more than a hundred speakers. We used the datasets to design and evaluate a novel self-supervised audio representation that leverages the effectiveness of handcrafted features (DSP-based) and the complexity of data-driven DNN representations. Notably, the proposed approach outperformed both extensive handcrafted feature sets and novel DNN-based audio representation learning approaches., Comment: Submitted to InterSpeech 2022

This paper presents AC-VC (Almost Causal Voice Conversion), a phonetic posteriorgrams based voice conversion system that can perform any-to-many voice conversion while having only 57.5 ms future look-ahead. The complete system is composed of three neural networks trained separately with non-parallel data. While most of the current voice conversion systems focus primarily on quality irrespective of algorithmic latency, this work elaborates on designing a method using a minimal amount of future context thus allowing a future real-time implementation. According to a subjective listening test organized in this work, the proposed AC-VC system achieves parity with the non-causal ASR-TTS baseline of the Voice Conversion Challenge 2020 in naturalness with a MOS of 3.5. In contrast, the results indicate that missing future context impacts speaker similarity. Obtained similarity percentage of 65% is lower than the similarity of current best voice conversion systems., Comment: ASRU 2021

At the end of Moore's law, new computing paradigms are required to prolong the battery life of wearable and IoT smart audio devices. Theoretical analysis and physical validation have shown that analog signal processing (ASP) can be more power-efficient than its digital counterpart in the realm of low-to-medium signal-to-noise ratio applications. In addition, ASP allows a direct interface with an analog microphone without a power-hungry analog-to-digital converter. Here, we present power-efficient analog acoustic features (PEAF) that are validated by fabricated CMOS chips for running audio recognition. Linear, non-linear, and learnable PEAF variants are evaluated on two speech processing tasks that are demanded in many battery-operated devices: wake word detection (WWD) and keyword spotting (KWS). Compared to digital acoustic features, higher power efficiency with competitive classification accuracy can be obtained. A novel theoretical framework based on information theory is established to analyze the information flow in each individual stage of the feature extraction pipeline. The analysis identifies the information bottleneck and helps improve the KWS accuracy by up to 7%. This work may pave the way to building more power-efficient smart audio devices with best-in-class inference performance., Comment: Submitted to Interspeech 2022

Recent developments in speech emotion recognition (SER) often leverage deep neural networks (DNNs). Comparing and benchmarking different DNN models can often be tedious due to the use of different datasets and evaluation protocols. To facilitate the process, here, we present the Speech Emotion Recognition Adaptation Benchmark (SERAB), a framework for evaluating the performance and generalization capacity of different approaches for utterance-level SER. The benchmark is composed of nine datasets for SER in six languages. Since the datasets have different sizes and numbers of emotional classes, the proposed setup is particularly suitable for estimating the generalization capacity of pre-trained DNN-based feature extractors. We used the proposed framework to evaluate a selection of standard hand-crafted feature sets and state-of-the-art DNN representations. The results highlight that using only a subset of the data included in SERAB can result in biased evaluation, while compliance with the proposed protocol can circumvent this issue., Comment: Submitted to ICASSP 2022

Speech audio quality is subject to degradation caused by an acoustic environment and isotropic ambient and point noises. The environment can lead to decreased speech intelligibility and loss of focus and attention by the listener. Basic acoustic parameters that characterize the environment well are (i) signal-to-noise ratio (SNR), (ii) speech transmission index, (iii) reverberation time, (iv) clarity, and (v) direct-to-reverberant ratio. Except for the SNR, these parameters are usually derived from the Room Impulse Response (RIR) measurements; however, such measurements are often not available. This work presents a universal room acoustic estimator design based on convolutional recurrent neural networks that estimate the acoustic environment measurement blindly and jointly. Our results indicate that the proposed system is robust to non-stationary signal variations and outperforms current state-of-the-art methods., Comment: Room acoustics, Convolutional Recurrent Neural Network, RT60, C50, DRR, STI, SNR

Acoustic environment characterization opens doors for sound reproduction innovations, smart EQing, speech enhancement, hearing aids, and forensics. Reverberation time, clarity, and direct-to-reverberant ratio are acoustic parameters that have been defined to describe reverberant environments. They are closely related to speech intelligibility and sound quality. As explained in the ISO3382 standard, they can be derived from a room measurement called the Room Impulse Response (RIR). However, measuring RIRs requires specific equipment and intrusive sound to be played. The recent audio combined with machine learning suggests that one could estimate those parameters blindly using speech or music signals. We follow these advances and propose a robust end-to-end method to achieve blind joint acoustic parameter estimation using speech and/or music signals. Our results indicate that convolutional recurrent neural networks perform best for this task, and including music in training also helps improve inference from speech.

Music source separation represents the task of extracting all the instruments from a given song. Recent breakthroughs on this challenge have gravitated around a single dataset, MUSDB, only limited to four instrument classes. Larger datasets and more instruments are costly and time-consuming in collecting data and training deep neural networks (DNNs). In this work, we propose a fast method to evaluate the separability of instruments in any dataset without training and tuning a DNN. This separability measure helps to select appropriate samples for the efficient training of neural networks. Based on the oracle principle with an ideal ratio mask, our approach is an excellent proxy to estimate the separation performances of state-of-the-art deep learning approaches such as TasNet or Open-Unmix. Our results contribute to revealing two essential points for audio source separation: 1) the ideal ratio mask, although light and straightforward, provides an accurate measure of the audio separability performance of recent neural nets, and 2) new end-to-end learning methods such as Tasnet, that operate directly on waveforms, are, in fact, internally building a Time-Frequency (TF) representation, so that they encounter the same limitations as the TF based-methods when separating audio pattern overlapping in the TF plane.

This paper introduces FastVC, an end-to-end model for fast Voice Conversion (VC). The proposed model can convert speech of arbitrary length from multiple source speakers to multiple target speakers. FastVC is based on a conditional AutoEncoder (AE) trained on non-parallel data and requires no annotations at all. This model's latent representation is shown to be speaker-independent and similar to phonemes, which is a desirable feature for VC systems. While the current VC systems primarily focus on achieving the highest overall speech quality, this paper tries to balance the development concerning resources needed to run the systems. Despite the simple structure of the proposed model, it outperforms the VC Challenge 2020 baselines on the cross-lingual task in terms of naturalness.

Advances of deep learning for Artificial Neural Networks(ANNs) have led to significant improvements in the performance of digital signal processing systems implemented on digital chips. Although recent progress in low-power chips is remarkable, neuromorphic chips that run Spiking Neural Networks (SNNs) based applications offer an even lower power consumption, as a consequence of the ensuing sparse spike-based coding scheme. In this work, we develop a SNN-based Voice Activity Detection (VAD) system that belongs to the building blocks of any audio and speech processing system. We propose to use the bin encoding, a novel method to convert log mel filterbank bins of single-time frames into spike patterns. We integrate the proposed scheme in a bilayer spiking architecture which was evaluated on the QUT-NOISE-TIMIT corpus. Our approach shows that SNNs enable an ultra low-power implementation of a VAD classifier that consumes only 3.8$\mu$W, while achieving state-of-the-art performance., Comment: 5 pages, 3 figures, 1 table

Recent advances in Voice Activity Detection (VAD) are driven by artificial and Recurrent Neural Networks (RNNs), however, using a VAD system in battery-operated devices requires further power efficiency. This can be achieved by neuromorphic hardware, which enables Spiking Neural Networks (SNNs) to perform inference at very low energy consumption. Spiking networks are characterized by their ability to process information efficiently, in a sparse cascade of binary events in time called spikes. However, a big performance gap separates artificial from spiking networks, mostly due to a lack of powerful SNN training algorithms. To overcome this problem we exploit an SNN model that can be recast into an RNN-like model and trained with known deep learning techniques. We describe an SNN training procedure that achieves low spiking activity and pruning algorithms to remove 85% of the network connections with no performance loss. The model achieves state-of-the-art performance with a fraction of power consumption comparing to other methods., Comment: 5 pages, 2 figures, 2 tables

Recent breakthroughs in deep learning often rely on representation learning and knowledge transfer. In recent years, unsupervised and self-supervised techniques for learning speech representation were developed to foster automatic speech recognition. Up to date, most of these approaches are task-specific and designed for within-task transfer learning between different datasets or setups of a particular task. In turn, learning task-independent representation of speech and cross-task applications of transfer learning remain less common. Here, we introduce an encoder capturing word-level representations of speech for cross-task transfer learning. We demonstrate the application of the pre-trained encoder in four distinct speech and audio processing tasks: (i) speech enhancement, (ii) language identification, (iii) speech, noise, and music classification, and (iv) speaker identification. In each task, we compare the performance of our cross-task transfer learning approach to task-specific baselines. Our results show that the speech representation captured by the encoder through the pre-training is transferable across distinct speech processing tasks and datasets. Notably, even simple applications of our pre-trained encoder outperformed task-specific methods, or were comparable, depending on the task., Comment: Published at EUSIPCO 2021

Transient loud intrusions, often occurring in noisy environments, can completely overpower speech signal and lead to an inevitable loss of information. While existing algorithms for noise suppression can yield impressive results, their efficacy remains limited for very low signal-to-noise ratios or when parts of the signal are missing. To address these limitations, here we propose an end-to-end framework for speech inpainting, the context-based retrieval of missing or severely distorted parts of time-frequency representation of speech. The framework is based on a convolutional U-Net trained via deep feature losses, obtained using speechVGG, a deep speech feature extractor pre-trained on an auxiliary word classification task. Our evaluation results demonstrate that the proposed framework can recover large portions of missing or distorted time-frequency representation of speech, up to 400 ms and 3.2 kHz in bandwidth. In particular, our approach provided a substantial increase in STOI & PESQ objective metrics of the initially corrupted speech samples. Notably, using deep feature losses to train the framework led to the best results, as compared to conventional approaches., Comment: Accepted to InterSpeech2020

Most current very low bit rate (VLBR) speech coding systems use hidden Markov model (HMM) based speech recognition/synthesis techniques. This allows transmission of information (such as phonemes) segment by segment that decreases the bit rate. However, the encoder based on a phoneme speech recognition may create bursts of segmental errors. Segmental errors are further propagated to optional suprasegmental (such as syllable) information coding. Together with the errors of voicing detection in pitch parametrization, HMM-based speech coding creates speech discontinuities and unnatural speech sound artefacts. In this paper, we propose a novel VLBR speech coding framework based on neural networks (NNs) for end-to-end speech analysis and synthesis without HMMs. The speech coding framework relies on phonological (sub-phonetic) representation of speech, and it is designed as a composition of deep and spiking NNs: a bank of phonological analysers at the transmitter, and a phonological synthesizer at the receiver, both realised as deep NNs, and a spiking NN as an incremental and robust encoder of syllable boundaries for coding of continuous fundamental frequency (F0). A combination of phonological features defines much more sound patterns than phonetic features defined by HMM-based speech coders, and the finer analysis/synthesis code contributes into smoother encoded speech. Listeners significantly prefer the NN-based approach due to fewer discontinuities and speech artefacts of the encoded speech. A single forward pass is required during the speech encoding and decoding. The proposed VLBR speech coding operates at a bit rate of approximately 360 bits/s.

Using phonological speech vocoding, we propose a platform for exploring relations between phonology and speech processing, and in broader terms, for exploring relations between the abstract and physical structures of a speech signal. Our goal is to make a step towards bridging phonology and speech processing and to contribute to the program of Laboratory Phonology. We show three application examples for laboratory phonology: compositional phonological speech modelling, a comparison of phonological systems and an experimental phonological parametric text-to-speech (TTS) system. The featural representations of the following three phonological systems are considered in this work: (i) Government Phonology (GP), (ii) the Sound Pattern of English (SPE), and (iii) the extended SPE (eSPE). Comparing GP- and eSPE-based vocoded speech, we conclude that the latter achieves slightly better results than the former. However, GP - the most compact phonological speech representation - performs comparably to the systems with a higher number of phonological features. The parametric TTS based on phonological speech representation, and trained from an unlabelled audiobook in an unsupervised manner, achieves intelligibility of 85% of the state-of-the-art parametric speech synthesis. We envision that the presented approach paves the way for researchers in both fields to form meaningful hypotheses that are explicitly testable using the concepts developed and exemplified in this paper. On the one hand, laboratory phonologists might test the applied concepts of their theoretical models, and on the other hand, the speech processing community may utilize the concepts developed for the theoretical phonological models for improvements of the current state-of-the-art applications.

The speech signal conveys information on different time scales from short time scale or segmental, associated to phonological and phonetic information to long time scale or supra segmental, associated to syllabic and prosodic information. Linguistic and neurocognitive studies recognize the phonological classes at segmental level as the essential and invariant representations used in speech temporal organization. In the context of speech processing, a deep neural network (DNN) is an effective computational method to infer the probability of individual phonological classes from a short segment of speech signal. A vector of all phonological class probabilities is referred to as phonological posterior. There are only very few classes comprising a short term speech signal; hence, the phonological posterior is a sparse vector. Although the phonological posteriors are estimated at segmental level, we claim that they convey supra-segmental information. Specifically, we demonstrate that phonological posteriors are indicative of syllabic and prosodic events. Building on findings from converging linguistic evidence on the gestural model of Articulatory Phonology as well as the neural basis of speech perception, we hypothesize that phonological posteriors convey properties of linguistic classes at multiple time scales, and this information is embedded in their support (index) of active coefficients. To verify this hypothesis, we obtain a binary representation of phonological posteriors at the segmental level which is referred to as first-order sparsity structure; the high-order structures are obtained by the concatenation of first-order binary vectors. It is then confirmed that the classification of supra-segmental linguistic events, the problem known as linguistic parsing, can be achieved with high accuracy using asimple binary pattern matching of first-order or high-order structures.

Human laugh is able to convey various kinds of meanings in human communications. There exists various kinds of human laugh signal, for example: vocalized laugh and non vocalized laugh. Following the theories of psychology, among all the vocalized laugh type, rhythmic staccato-vocalization significantly evokes the positive responses in the interactions. In this paper we attempt to exploit this observation to detect human laugh occurrences, i.e., the laughter, in multiparty conversations from the AMI meeting corpus. First, we separate the high energy frames from speech, leaving out the low energy frames through power spectral density estimation. We borrow the algorithm of rhythm detection from the area of music analysis to use that on the high energy frames. Finally, we detect rhythmic laugh frames, analyzing the candidate rhythmic frames using statistics. This novel approach for detection of `positive' rhythmic human laughter performs better than the standard laughter classification baseline., Comment: 5 pages, 1 figure, conference paper