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.
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.
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.
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.
The zero-shot text-to-speech (TTS) method, based on speaker embeddings extracted from reference speech using self-supervised learning (SSL) speech representations, can reproduce speaker characteristics very accurately. However, this approach suffers from degradation in speech synthesis quality when the reference speech contains noise. In this paper, we propose a noise-robust zero-shot TTS method. We incorporated adapters into the SSL model, which we fine-tuned with the TTS model using noisy reference speech. In addition, to further improve performance, we adopted a speech enhancement (SE) front-end. With these improvements, our proposed SSL-based zero-shot TTS achieved high-quality speech synthesis with noisy reference speech. Through the objective and subjective evaluations, we confirmed that the proposed method is highly robust to noise in reference speech, and effectively works in combination with SE.
Confidence estimation, in which we estimate the reliability of each recognized token (e.g., word, sub-word, and character) in automatic speech recognition (ASR) hypotheses and detect incorrectly recognized tokens, is an important function for developing ASR applications. In this study, we perform confidence estimation for end-to-end (E2E) ASR hypotheses. Recent E2E ASR systems show high performance (e.g., around 5% token error rates) for various ASR tasks. In such situations, confidence estimation becomes difficult since we need to detect infrequent incorrect tokens from mostly correct token sequences. To tackle this imbalanced dataset problem, we employ a bidirectional long short-term memory (BLSTM)-based model as a strong binary-class (correct/incorrect) sequence labeler that is trained with a class balancing objective. We experimentally confirmed that, by utilizing several types of ASR decoding scores as its auxiliary features, the model steadily shows high confidence estimation performance under highly imbalanced settings. We also confirmed that the BLSTM-based model outperforms Transformer-based confidence estimation models, which greatly underestimate incorrect tokens.
We investigate the effectiveness of using a large ensemble of advanced neural language models (NLMs) for lattice rescoring on automatic speech recognition (ASR) hypotheses. Previous studies have reported the effectiveness of combining a small number of NLMs. In contrast, in this study, we combine up to eight NLMs, i.e., forward/backward long short-term memory/Transformer-LMs that are trained with two different random initialization seeds. We combine these NLMs through iterative lattice generation. Since these NLMs work complementarily with each other, by combining them one by one at each rescoring iteration, language scores attached to given lattice arcs can be gradually refined. Consequently, errors of the ASR hypotheses can be gradually reduced. We also investigate the effectiveness of carrying over contextual information (previous rescoring results) across a lattice sequence of a long speech such as a lecture speech. In experiments using a lecture speech corpus, by combining the eight NLMs and using context carry-over, we obtained a 24.4% relative word error rate reduction from the ASR 1-best baseline. For further comparison, we performed simultaneous (i.e., non-iterative) NLM combination and 100-best rescoring using the large ensemble of NLMs, which confirmed the advantage of lattice rescoring with iterative NLM combination.
Jointly training a speech enhancement (SE) front-end and an automatic speech recognition (ASR) back-end has been investigated as a way to mitigate the influence of \emph{processing distortion} generated by single-channel SE on ASR. In this paper, we investigate the effect of such joint training on the signal-level characteristics of the enhanced signals from the viewpoint of the decomposed noise and artifact errors. The experimental analyses provide two novel findings: 1) ASR-level training of the SE front-end reduces the artifact errors while increasing the noise errors, and 2) simply interpolating the enhanced and observed signals, which achieves a similar effect of reducing artifacts and increasing noise, improves ASR performance without jointly modifying the SE and ASR modules, even for a strong ASR back-end using a WavLM feature extractor. Our findings provide a better understanding of the effect of joint training and a novel insight for designing an ASR agnostic SE front-end.
Array processing performance depends on the number of microphones available. Virtual microphone estimation (VME) has been proposed to increase the number of microphone signals artificially. Neural network-based VME (NN-VME) trains an NN with a VM-level loss to predict a signal at a microphone location that is available during training but not at inference. However, this training objective may not be optimal for a specific array processing back-end, such as beamforming. An alternative approach is to use a training objective considering the array-processing back-end, such as a loss on the beamformer output. This approach may generate signals optimal for beamforming but not physically grounded. To combine the advantages of both approaches, this paper proposes a multi-task loss for NN-VME that combines both VM-level and beamformer-level losses. We evaluate the proposed multi-task NN-VME on multi-talker underdetermined conditions and show that it achieves a 33.1 % relative WER improvement compared to using only real microphones and 10.8 % compared to using a prior NN-VME approach.
We propose a new shallow fusion (SF) method to exploit an external backward language model (BLM) for end-to-end automatic speech recognition (ASR). The BLM has complementary characteristics with a forward language model (FLM), and the effectiveness of their combination has been confirmed by rescoring ASR hypotheses as post-processing. In the proposed SF, we iteratively apply the BLM to partial ASR hypotheses in the backward direction (i.e., from the possible next token to the start symbol) during decoding, substituting the newly calculated BLM scores for the scores calculated at the last iteration. To enhance the effectiveness of this iterative SF (ISF), we train a partial sentence-aware BLM (PBLM) using reversed text data including partial sentences, considering the framework of ISF. In experiments using an attention-based encoder-decoder ASR system, we confirmed that ISF using the PBLM shows comparable performance with SF using the FLM. By performing ISF, early pruning of prospective hypotheses can be prevented during decoding, and we can obtain a performance improvement compared to applying the PBLM as post-processing. Finally, we confirmed that, by combining SF and ISF, further performance improvement can be obtained thanks to the complementarity of the FLM and PBLM.