Interaural time difference loss for binaural target sound extraction

Binaural target sound extraction (TSE) aims to extract a desired sound from a binaural mixture of arbitrary sounds while preserving the spatial cues of the desired sound. Indeed, for many applications, the target sound signal and its spatial cues carry important information about the sound source. Binaural TSE can be realized with a neural network trained to output only the desired sound given a binaural mixture and an embedding characterizing the desired sound class as inputs. Conventional TSE systems are trained using signal-level losses, which measure the difference between the extracted and reference signals for the left and right channels. In this paper, we propose adding explicit spatial losses to better preserve the spatial cues of the target sound. In particular, we explore losses aiming at preserving the interaural level (ILD), phase (IPD), and time differences (ITD). We show experimentally that adding such spatial losses, particularly our newly proposed ITD loss, helps preserve better spatial cues while maintaining the signal-level metrics.

Sentence-wise Speech Summarization: Task, Datasets, and End-to-End Modeling with LM Knowledge Distillation

This paper introduces a novel approach called sentence-wise speech summarization (Sen-SSum), which generates text summaries from a spoken document in a sentence-by-sentence manner. Sen-SSum combines the real-time processing of automatic speech recognition (ASR) with the conciseness of speech summarization. To explore this approach, we present two datasets for Sen-SSum: Mega-SSum and CSJ-SSum. Using these datasets, our study evaluates two types of Transformer-based models: 1) cascade models that combine ASR and strong text summarization models, and 2) end-to-end (E2E) models that directly convert speech into a text summary. While E2E models are appealing to develop compute-efficient models, they perform worse than cascade models. Therefore, we propose knowledge distillation for E2E models using pseudo-summaries generated by the cascade models. Our experiments show that this proposed knowledge distillation effectively improves the performance of the E2E model on both datasets.

Lightweight Zero-shot Text-to-Speech with Mixture of Adapters

The advancements in zero-shot text-to-speech (TTS) methods, based on large-scale models, have demonstrated high fidelity in reproducing speaker characteristics. However, these models are too large for practical daily use. We propose a lightweight zero-shot TTS method using a mixture of adapters (MoA). Our proposed method incorporates MoA modules into the decoder and the variance adapter of a non-autoregressive TTS model. These modules enhance the ability to adapt a wide variety of speakers in a zero-shot manner by selecting appropriate adapters associated with speaker characteristics on the basis of speaker embeddings. Our method achieves high-quality speech synthesis with minimal additional parameters. Through objective and subjective evaluations, we confirmed that our method achieves better performance than the baseline with less than 40\% of parameters at 1.9 times faster inference speed. Audio samples are available on our demo page (https://ntt-hilab-gensp.github.io/is2024lightweightTTS/).

SpeakerBeam-SS: Real-time Target Speaker Extraction with Lightweight Conv-TasNet and State Space Modeling

Real-time target speaker extraction (TSE) is intended to extract the desired speaker's voice from the observed mixture of multiple speakers in a streaming manner. Implementing real-time TSE is challenging as the computational complexity must be reduced to provide real-time operation. This work introduces to Conv-TasNet-based TSE a new architecture based on state space modeling (SSM) that has been shown to model long-term dependency effectively. Owing to SSM, fewer dilated convolutional layers are required to capture temporal dependency in Conv-TasNet, resulting in the reduction of model complexity. We also enlarge the window length and shift of the convolutional (TasNet) frontend encoder to reduce the computational cost further; the performance decline is compensated by over-parameterization of the frontend encoder. The proposed method reduces the real-time factor by 78% from the conventional causal Conv-TasNet-based TSE while matching its performance.

Applying LLMs for Rescoring N-best ASR Hypotheses of Casual Conversations: Effects of Domain Adaptation and Context Carry-over

Large language models (LLMs) have been successfully applied for rescoring automatic speech recognition (ASR) hypotheses. However, their ability to rescore ASR hypotheses of casual conversations has not been sufficiently explored. In this study, we reveal it by performing N-best ASR hypotheses rescoring using Llama2 on the CHiME-7 distant ASR (DASR) task. Llama2 is one of the most representative LLMs, and the CHiME-7 DASR task provides datasets of casual conversations between multiple participants. We investigate the effects of domain adaptation of the LLM and context carry-over when performing N-best rescoring. Experimental results show that, even without domain adaptation, Llama2 outperforms a standard-size domain-adapted Transformer-LM, especially when using a long context. Domain adaptation shortens the context length needed with Llama2 to achieve its best performance, i.e., it reduces the computational cost of Llama2.

Rethinking Processing Distortions: Disentangling the Impact of Speech Enhancement Errors on Speech Recognition Performance

It is challenging to improve automatic speech recognition (ASR) performance in noisy conditions with a single-channel speech enhancement (SE) front-end. This is generally attributed to the processing distortions caused by the nonlinear processing of single-channel SE front-ends. However, the causes of such degraded ASR performance have not been fully investigated. How to design single-channel SE front-ends in a way that significantly improves ASR performance remains an open research question. In this study, we investigate a signal-level numerical metric that can explain the cause of degradation in ASR performance. To this end, we propose a novel analysis scheme based on the orthogonal projection-based decomposition of SE errors. This scheme manually modifies the ratio of the decomposed interference, noise, and artifact errors, and it enables us to directly evaluate the impact of each error type on ASR performance. Our analysis reveals the particularly detrimental effect of artifact errors on ASR performance compared to the other types of errors. This provides us with a more principled definition of processing distortions that cause the ASR performance degradation. Then, we study two practical approaches for reducing the impact of artifact errors. First, we prove that the simple observation adding (OA) post-processing (i.e., interpolating the enhanced and observed signals) can monotonically improve the signal-to-artifact ratio. Second, we propose a novel training objective, called artifact-boosted signal-to-distortion ratio (AB-SDR), which forces the model to estimate the enhanced signals with fewer artifact errors. Through experiments, we confirm that both the OA and AB-SDR approaches are effective in decreasing artifact errors caused by single-channel SE front-ends, allowing them to significantly improve ASR performance.

Target Speech Extraction with Pre-trained Self-supervised Learning Models

Pre-trained self-supervised learning (SSL) models have achieved remarkable success in various speech tasks. However, their potential in target speech extraction (TSE) has not been fully exploited. TSE aims to extract the speech of a target speaker in a mixture guided by enrollment utterances. We exploit pre-trained SSL models for two purposes within a TSE framework, i.e., to process the input mixture and to derive speaker embeddings from the enrollment. In this paper, we focus on how to effectively use SSL models for TSE. We first introduce a novel TSE downstream task following the SUPERB principles. This simple experiment shows the potential of SSL models for TSE, but extraction performance remains far behind the state-of-the-art. We then extend a powerful TSE architecture by incorporating two SSL-based modules: an Adaptive Input Enhancer (AIE) and a speaker encoder. Specifically, the proposed AIE utilizes intermediate representations from the CNN encoder by adjusting the time resolution of CNN encoder and transformer blocks through progressive upsampling, capturing both fine-grained and hierarchical features. Our method outperforms current TSE systems achieving a SI-SDR improvement of 14.0 dB on LibriMix. Moreover, we can further improve performance by 0.7 dB by fine-tuning the whole model including the SSL model parameters.

Probing Self-supervised Learning Models with Target Speech Extraction

Large-scale pre-trained self-supervised learning (SSL) models have shown remarkable advancements in speech-related tasks. However, the utilization of these models in complex multi-talker scenarios, such as extracting a target speaker in a mixture, is yet to be fully evaluated. In this paper, we introduce target speech extraction (TSE) as a novel downstream task to evaluate the feature extraction capabilities of pre-trained SSL models. TSE uniquely requires both speaker identification and speech separation, distinguishing it from other tasks in the Speech processing Universal PERformance Benchmark (SUPERB) evaluation. Specifically, we propose a TSE downstream model composed of two lightweight task-oriented modules based on the same frozen SSL model. One module functions as a speaker encoder to obtain target speaker information from an enrollment speech, while the other estimates the target speaker's mask to extract its speech from the mixture. Experimental results on the Libri2mix datasets reveal the relevance of the TSE downstream task to probe SSL models, as its performance cannot be simply deduced from other related tasks such as speaker verification and separation.

Array Geometry-Robust Attention-Based Neural Beamformer for Moving Speakers

Recently, a mask-based beamformer with attention-based spatial covariance matrix aggregator (ASA) was proposed, which was demonstrated to track moving sources accurately. However, the deep neural network model used in this algorithm is limited to a specific channel configuration, requiring a different model in case a different channel permutation, channel count, or microphone array geometry is considered. Addressing this limitation, in this paper, we investigate three approaches to improve the robustness of the ASA-based tracking method against such variations: incorporating random channel configurations during the training process, employing the transform-average-concatenate (TAC) method to process multi-channel input features (allowing for any channel count and enabling permutation invariance), and utilizing input features that are robust against variations of the channel configuration. Our experiments, conducted using the CHiME-3 and DEMAND datasets, demonstrate improved robustness against mismatches in channel permutations, channel counts, and microphone array geometries compared to the conventional ASA-based tracking method without compromising performance in matched conditions, suggesting that the mask-based beamformer with ASA integrating the proposed approaches has the potential to track moving sources for arbitrary microphone arrays.

What Do Self-Supervised Speech and Speaker Models Learn? New Findings From a Cross Model Layer-Wise Analysis

Self-supervised learning (SSL) has attracted increased attention for learning meaningful speech representations. Speech SSL models, such as WavLM, employ masked prediction training to encode general-purpose representations. In contrast, speaker SSL models, exemplified by DINO-based models, adopt utterance-level training objectives primarily for speaker representation. Understanding how these models represent information is essential for refining model efficiency and effectiveness. Unlike the various analyses of speech SSL, there has been limited investigation into what information speaker SSL captures and how its representation differs from speech SSL or other fully-supervised speaker models. This paper addresses these fundamental questions. We explore the capacity to capture various speech properties by applying SUPERB evaluation probing tasks to speech and speaker SSL models. We also examine which layers are predominantly utilized for each task to identify differences in how speech is represented. Furthermore, we conduct direct comparisons to measure the similarities between layers within and across models. Our analysis unveils that 1) the capacity to represent content information is somewhat unrelated to enhanced speaker representation, 2) specific layers of speech SSL models would be partly specialized in capturing linguistic information, and 3) speaker SSL models tend to disregard linguistic information but exhibit more sophisticated speaker representation.