Tatsuya Komatsu

Tatsuya Komatsu, Yusuke Fujita, Jaesong Lee et al.

This paper proposes a method for improved CTC inference with searched intermediates and multi-pass conditioning. The paper first formulates self-conditioned CTC as a probabilistic model with an intermediate prediction as a latent representation and provides a tractable conditioning framework. We then propose two new conditioning methods based on the new formulation: (1) Searched intermediate conditioning that refines intermediate predictions with beam-search, (2) Multi-pass conditioning that uses predictions of previous inference for conditioning the next inference. These new approaches enable better conditioning than the original self-conditioned CTC during inference and improve the final performance. Experiments with the LibriSpeech dataset show relative 3%/12% performance improvement at the maximum in test clean/other sets compared to the original self-conditioned CTC.

Tatsuya Komatsu, Yusuke Fujita, Kazuya Takeda et al.

This study introduces a novel training paradigm, audio difference learning, for improving audio captioning. The fundamental concept of the proposed learning method is to create a feature representation space that preserves the relationship between audio, enabling the generation of captions that detail intricate audio information. This method employs a reference audio along with the input audio, both of which are transformed into feature representations via a shared encoder. Captions are then generated from these differential features to describe their differences. Furthermore, a unique technique is proposed that involves mixing the input audio with additional audio, and using the additional audio as a reference. This results in the difference between the mixed audio and the reference audio reverting back to the original input audio. This allows the original input's caption to be used as the caption for their difference, eliminating the need for additional annotations for the differences. In the experiments using the Clotho and ESC50 datasets, the proposed method demonstrated an improvement in the SPIDEr score by 7% compared to conventional methods.

Reo Shimizu, Ryuichi Yamamoto, Masaya Kawamura et al.

We propose PromptTTS++, a prompt-based text-to-speech (TTS) synthesis system that allows control over speaker identity using natural language descriptions. To control speaker identity within the prompt-based TTS framework, we introduce the concept of speaker prompt, which describes voice characteristics (e.g., gender-neutral, young, old, and muffled) designed to be approximately independent of speaking style. Since there is no large-scale dataset containing speaker prompts, we first construct a dataset based on the LibriTTS-R corpus with manually annotated speaker prompts. We then employ a diffusion-based acoustic model with mixture density networks to model diverse speaker factors in the training data. Unlike previous studies that rely on style prompts describing only a limited aspect of speaker individuality, such as pitch, speaking speed, and energy, our method utilizes an additional speaker prompt to effectively learn the mapping from natural language descriptions to the acoustic features of diverse speakers. Our subjective evaluation results show that the proposed method can better control speaker characteristics than the methods without the speaker prompt. Audio samples are available at https://reppy4620.github.io/demo.promptttspp/.

Yusuke Fujita, Tatsuya Komatsu, Robin Scheibler et al.

End-to-end neural diarization (EEND) with encoder-decoder-based attractors (EDA) is a promising method to handle the whole speaker diarization problem simultaneously with a single neural network. While the EEND model can produce all frame-level speaker labels simultaneously, it disregards output label dependency. In this work, we propose a novel EEND model that introduces the label dependency between frames. The proposed method generates non-autoregressive intermediate attractors to produce speaker labels at the lower layers and conditions the subsequent layers with these labels. While the proposed model works in a non-autoregressive manner, the speaker labels are refined by referring to the whole sequence of intermediate labels. The experiments with the two-speaker CALLHOME dataset show that the intermediate labels with the proposed non-autoregressive intermediate attractors boost the diarization performance. The proposed method with the deeper network benefits more from the intermediate labels, resulting in better performance and training throughput than EEND-EDA.

Taichi Nishimura, Shota Nakada, Hokuto Munakata et al.

We propose Lighthouse, a user-friendly library for reproducible video moment retrieval and highlight detection (MR-HD). Although researchers proposed various MR-HD approaches, the research community holds two main issues. The first is a lack of comprehensive and reproducible experiments across various methods, datasets, and video-text features. This is because no unified training and evaluation codebase covers multiple settings. The second is user-unfriendly design. Because previous works use different libraries, researchers set up individual environments. In addition, most works release only the training codes, requiring users to implement the whole inference process of MR-HD. Lighthouse addresses these issues by implementing a unified reproducible codebase that includes six models, three features, and five datasets. In addition, it provides an inference API and web demo to make these methods easily accessible for researchers and developers. Our experiments demonstrate that Lighthouse generally reproduces the reported scores in the reference papers. The code is available at https://github.com/line/lighthouse.

Yu Nakagome, Tatsuya Komatsu, Yusuke Fujita et al.

This paper proposes InterAug: a novel training method for CTC-based ASR using augmented intermediate representations for conditioning. The proposed method exploits the conditioning framework of self-conditioned CTC to train robust models by conditioning with "noisy" intermediate predictions. During the training, intermediate predictions are changed to incorrect intermediate predictions, and fed into the next layer for conditioning. The subsequent layers are trained to correct the incorrect intermediate predictions with the intermediate losses. By repeating the augmentation and the correction, iterative refinements, which generally require a special decoder, can be realized only with the audio encoder. To produce noisy intermediate predictions, we also introduce new augmentation: intermediate feature space augmentation and intermediate token space augmentation that are designed to simulate typical errors. The combination of the proposed InterAug framework with new augmentation allows explicit training of the robust audio encoders. In experiments using augmentations simulating deletion, insertion, and substitution error, we confirmed that the trained model acquires robustness to each error, boosting the speech recognition performance of the strong self-conditioned CTC baseline.

Yusuke Fujita, Tatsuya Komatsu, Yusuke Kida

End-to-end automatic speech recognition directly maps input speech to characters. However, the mapping can be problematic when several different pronunciations should be mapped into one character or when one pronunciation is shared among many different characters. Japanese ASR suffers the most from such many-to-one and one-to-many mapping problems due to Japanese kanji characters. To alleviate the problems, we introduce explicit interaction between characters and syllables using Self-conditioned connectionist temporal classification (CTC), in which the upper layers are ``self-conditioned'' on the intermediate predictions from the lower layers. The proposed method utilizes character-level and syllable-level intermediate predictions as conditioning features to deal with mutual dependency between characters and syllables. Experimental results on Corpus of Spontaneous Japanese show that the proposed method outperformed the conventional multi-task and Self-conditioned CTC methods.

Hokuto Munakata, Taichi Nishimura, Shota Nakada et al.

In this paper, we propose and design a new task called audio moment retrieval (AMR). Unlike conventional language-based audio retrieval tasks that search for short audio clips from an audio database, AMR aims to predict relevant moments in untrimmed long audio based on a text query. Given the lack of prior work in AMR, we first build a dedicated dataset, Clotho-Moment, consisting of large-scale simulated audio recordings with moment annotations. We then propose a DETR-based model, named Audio Moment DETR (AM-DETR), as a fundamental framework for AMR tasks. This model captures temporal dependencies within audio features, inspired by similar video moment retrieval tasks, thus surpassing conventional clip-level audio retrieval methods. Additionally, we provide manually annotated datasets to properly measure the effectiveness and robustness of our methods on real data. Experimental results show that AM-DETR, trained with Clotho-Moment, outperforms a baseline model that applies a clip-level audio retrieval method with a sliding window on all metrics, particularly improving Recall1@0.7 by 9.00 points. Our datasets and code are publicly available in https://h-munakata.github.io/Language-based-Audio-Moment-Retrieval.

Shota Nakada, Taichi Nishimura, Hokuto Munakata et al.

Current audio-visual representation learning can capture rough object categories (e.g., ``animals'' and ``instruments''), but it lacks the ability to recognize fine-grained details, such as specific categories like ``dogs'' and ``flutes'' within animals and instruments. To address this issue, we introduce DETECLAP, a method to enhance audio-visual representation learning with object information. Our key idea is to introduce an audio-visual label prediction loss to the existing Contrastive Audio-Visual Masked AutoEncoder to enhance its object awareness. To avoid costly manual annotations, we prepare object labels from both audio and visual inputs using state-of-the-art language-audio models and object detectors. We evaluate the method of audio-visual retrieval and classification using the VGGSound and AudioSet20K datasets. Our method achieves improvements in recall@10 of +1.5% and +1.2% for audio-to-visual and visual-to-audio retrieval, respectively, and an improvement in accuracy of +0.6% for audio-visual classification.

Michael Hentschel, Yuta Nishikawa, Tatsuya Komatsu et al.

This study presents a novel approach for knowledge distillation (KD) from a BERT teacher model to an automatic speech recognition (ASR) model using intermediate layers. To distil the teacher's knowledge, we use an attention decoder that learns from BERT's token probabilities. Our method shows that language model (LM) information can be more effectively distilled into an ASR model using both the intermediate layers and the final layer. By using the intermediate layers as distillation target, we can more effectively distil LM knowledge into the lower network layers. Using our method, we achieve better recognition accuracy than with shallow fusion of an external LM, allowing us to maintain fast parallel decoding. Experiments on the LibriSpeech dataset demonstrate the effectiveness of our approach in enhancing greedy decoding with connectionist temporal classification (CTC).

Yuchi Ishikawa, Shota Nakada, Hokuto Munakata et al.

In this paper, we propose Language-Guided Contrastive Audio-Visual Masked Autoencoders (LG-CAV-MAE) to improve audio-visual representation learning. LG-CAV-MAE integrates a pretrained text encoder into contrastive audio-visual masked autoencoders, enabling the model to learn across audio, visual and text modalities. To train LG-CAV-MAE, we introduce an automatic method to generate audio-visual-text triplets from unlabeled videos. We first generate frame-level captions using an image captioning model and then apply CLAP-based filtering to ensure strong alignment between audio and captions. This approach yields high-quality audio-visual-text triplets without requiring manual annotations. We evaluate LG-CAV-MAE on audio-visual retrieval tasks, as well as an audio-visual classification task. Our method significantly outperforms existing approaches, achieving up to a 5.6% improvement in recall@10 for retrieval tasks and a 3.2% improvement for the classification task.

Hokuto Munakata, Takehiro Imamura, Taichi Nishimura et al.

We introduce CASTELLA, a human-annotated audio benchmark for the task of audio moment retrieval (AMR). Although AMR has various useful potential applications, there is still no established benchmark with real-world data. The early study of AMR trained the model with solely synthetic datasets. Moreover, the evaluation is based on annotated dataset of fewer than 100 samples. This resulted in less reliable reported performance. To ensure performance for applications in real-world environments, we present CASTELLA, a large-scale manually annotated AMR dataset. CASTELLA consists of 1,009, 213, and 640 audio recordings for train, valid, and test split, respectively, which is 24 times larger than the previous dataset. We also establish a baseline model for AMR using CASTELLA. Our experiments demonstrate that a model fine-tuned on CASTELLA after pre-training on the synthetic data outperformed a model trained solely on the synthetic data by 10.4 points in Recall1@0.7. CASTELLA is publicly available in https://h-munakata.github.io/CASTELLA-demo/.

Yuchi Ishikawa, Toranosuke Manabe, Tatsuya Komatsu et al.

Audio-visual captioning aims to generate holistic scene descriptions by jointly modeling sound and vision. While recent methods have improved performance through sophisticated modality fusion, it remains unclear to what extent the two modalities are complementary in current audio-visual captioning models and how robust these models are when one modality is degraded. We address these questions by conducting systematic modality robustness tests on LAVCap, a state-of-the-art audio-visual captioning model, in which we selectively suppress or corrupt the audio or visual streams to quantify sensitivity and complementarity. The analysis reveals a pronounced bias toward the audio stream in LAVCap. To evaluate how balanced audio-visual captioning models are in their use of both modalities, we augment AudioCaps with textual annotations that jointly describe the audio and visual streams, yielding the AudioVisualCaps dataset. In our experiments, we report LAVCap baseline results on AudioVisualCaps. We also evaluate the model under modality robustness tests on AudioVisualCaps and the results indicate that LAVCap trained on AudioVisualCaps exhibits less modality bias than when trained on AudioCaps.

Tatsuya Komatsu

This paper proposes CTC-based non-autoregressive ASR with self-conditioned folded encoders. The proposed method realizes non-autoregressive ASR with fewer parameters by folding the conventional stack of encoders into only two blocks; base encoders and folded encoders. The base encoders convert the input audio features into a neural representation suitable for recognition. This is followed by the folded encoders applied repeatedly for further refinement. Applying the CTC loss to the outputs of all encoders enforces the consistency of the input-output relationship. Thus, folded encoders learn to perform the same operations as an encoder with deeper distinct layers. In experiments, we investigate how to set the number of layers and the number of iterations for the base and folded encoders. The results show that the proposed method achieves a performance comparable to that of the conventional method using only 38% as many parameters. Furthermore, it outperforms the conventional method when increasing the number of iterations.

Tatsuya Komatsu, Shinji Watanabe, Koichi Miyazaki et al.

This paper proposes acoustic event detection (AED) with classifier chains, a new classifier based on the probabilistic chain rule. The proposed AED with classifier chains consists of a gated recurrent unit and performs iterative binary detection of each event one by one. In each iteration, the event's activity is estimated and used to condition the next output based on the probabilistic chain rule to form classifier chains. Therefore, the proposed method can handle the interdependence among events upon classification, while the conventional AED methods with multiple binary classifiers with a linear layer and sigmoid function have placed an assumption of conditional independence. In the experiments with a real-recording dataset, the proposed method demonstrates its superior AED performance to a relative 14.80% improvement compared to a convolutional recurrent neural network baseline system with the multiple binary classifiers.

Jin Sakuma, Tatsuya Komatsu, Robin Scheibler

We propose multi-layer perceptron (MLP)-based architectures suitable for variable length input. MLP-based architectures, recently proposed for image classification, can only be used for inputs of a fixed, pre-defined size. However, many types of data are naturally variable in length, for example, acoustic signals. We propose three approaches to extend MLP-based architectures for use with sequences of arbitrary length. The first one uses a circular convolution applied in the Fourier domain, the second applies a depthwise convolution, and the final relies on a shift operation. We evaluate the proposed architectures on an automatic speech recognition task with the Librispeech and Tedlium2 corpora. The best proposed MLP-based architectures improves WER by 1.0 / 0.9%, 0.9 / 0.5% on Librispeech dev-clean/dev-other, test-clean/test-other set, and 0.8 / 1.1% on Tedlium2 dev/test set using 86.4% the size of self-attention-based architecture.

Yosuke Higuchi, Nanxin Chen, Yuya Fujita et al.

Non-autoregressive (NAR) models simultaneously generate multiple outputs in a sequence, which significantly reduces the inference speed at the cost of accuracy drop compared to autoregressive baselines. Showing great potential for real-time applications, an increasing number of NAR models have been explored in different fields to mitigate the performance gap against AR models. In this work, we conduct a comparative study of various NAR modeling methods for end-to-end automatic speech recognition (ASR). Experiments are performed in the state-of-the-art setting using ESPnet. The results on various tasks provide interesting findings for developing an understanding of NAR ASR, such as the accuracy-speed trade-off and robustness against long-form utterances. We also show that the techniques can be combined for further improvement and applied to NAR end-to-end speech translation. All the implementations are publicly available to encourage further research in NAR speech processing.

Yusuke Kida, Tatsuya Komatsu, Masahito Togami

This paper proposes a novel label-synchronous speech-to-text alignment technique for automatic speech recognition (ASR). The speech-to-text alignment is a problem of splitting long audio recordings with un-aligned transcripts into utterance-wise pairs of speech and text. Unlike conventional methods based on frame-synchronous prediction, the proposed method re-defines the speech-to-text alignment as a label-synchronous text mapping problem. This enables an accurate alignment benefiting from the strong inference ability of the state-of-the-art attention-based encoder-decoder models, which cannot be applied to the conventional methods. Two different Transformer models named forward Transformer and backward Transformer are respectively used for estimating an initial and final tokens of a given speech segment based on end-of-sentence prediction with teacher-forcing. Experiments using the corpus of spontaneous Japanese (CSJ) demonstrate that the proposed method provides an accurate utterance-wise alignment, that matches the manually annotated alignment with as few as 0.2% errors. It is also confirmed that a Transformer-based hybrid CTC/Attention ASR model using the aligned speech and text pairs as an additional training data reduces character error rates relatively up to 59.0%, which is significantly better than 39.0% reduction by a conventional alignment method based on connectionist temporal classification model.

Jumon Nozaki, Tatsuya Komatsu

This paper proposes a method to relax the conditional independence assumption of connectionist temporal classification (CTC)-based automatic speech recognition (ASR) models. We train a CTC-based ASR model with auxiliary CTC losses in intermediate layers in addition to the original CTC loss in the last layer. During both training and inference, each generated prediction in the intermediate layers is summed to the input of the next layer to condition the prediction of the last layer on those intermediate predictions. Our method is easy to implement and retains the merits of CTC-based ASR: a simple model architecture and fast decoding speed. We conduct experiments on three different ASR corpora. Our proposed method improves a standard CTC model significantly (e.g., more than 20 % relative word error rate reduction on the WSJ corpus) with a little computational overhead. Moreover, for the TEDLIUM2 corpus and the AISHELL-1 corpus, it achieves a comparable performance to a strong autoregressive model with beam search, but the decoding speed is at least 30 times faster.

Tsubasa Takahashi, Shun Takagi, Hajime Ono et al.

This paper studies how to learn variational autoencoders with a variety of divergences under differential privacy constraints. We often build a VAE with an appropriate prior distribution to describe the desired properties of the learned representations and introduce a divergence as a regularization term to close the representations to the prior. Using differentially private SGD (DP-SGD), which randomizes a stochastic gradient by injecting a dedicated noise designed according to the gradient's sensitivity, we can easily build a differentially private model. However, we reveal that attaching several divergences increase the sensitivity from O(1) to O(B) in terms of batch size B. That results in injecting a vast amount of noise that makes it hard to learn. To solve the above issue, we propose term-wise DP-SGD that crafts randomized gradients in two different ways tailored to the compositions of the loss terms. The term-wise DP-SGD keeps the sensitivity at O(1) even when attaching the divergence. We can therefore reduce the amount of noise. In our experiments, we demonstrate that our method works well with two pairs of the prior distribution and the divergence.

Yoshiki Masuyama, Masahito Togami, Tatsuya Komatsu

This paper proposes a deep neural network (DNN)-based multi-channel speech enhancement system in which a DNN is trained to maximize the quality of the enhanced time-domain signal. DNN-based multi-channel speech enhancement is often conducted in the time-frequency (T-F) domain because spatial filtering can be efficiently implemented in the T-F domain. In such a case, ordinary objective functions are computed on the estimated T-F mask or spectrogram. However, the estimated spectrogram is often inconsistent, and its amplitude and phase may change when the spectrogram is converted back to the time-domain. That is, the objective function does not evaluate the enhanced time-domain signal properly. To address this problem, we propose to use an objective function defined on the reconstructed time-domain signal. Specifically, speech enhancement is conducted by multi-channel Wiener filtering in the T-F domain, and its result is converted back to the time-domain. We propose two objective functions computed on the reconstructed signal where the first one is defined in the time-domain, and the other one is defined in the T-F domain. Our experiment demonstrates the effectiveness of the proposed system comparing to T-F masking and mask-based beamforming.

Masahito Togami, Yoshiki Masuyama, Tatsuya Komatsu et al.

In this paper, we propose a multi-channel speech source separation with a deep neural network (DNN) which is trained under the condition that no clean signal is available. As an alternative to a clean signal, the proposed method adopts an estimated speech signal by an unsupervised speech source separation with a statistical model. As a statistical model of microphone input signal, we adopts a time-varying spatial covariance matrix (SCM) model which includes reverberation and background noise submodels so as to achieve robustness against reverberation and background noise. The DNN infers intermediate variables which are needed for constructing the time-varying SCM. Speech source separation is performed in a probabilistic manner so as to avoid overfitting to separation error. Since there are multiple intermediate variables, a loss function which evaluates a single intermediate variable is not applicable. Instead, the proposed method adopts a loss function which evaluates the output probabilistic signal directly based on Kullback-Leibler Divergence (KLD). Gradient of the loss function can be back-propagated into the DNN through all the intermediate variables. Experimental results under reverberant conditions show that the proposed method can train the DNN efficiently even when the number of training utterances is small, i.e., 1K.

Yuki Okamoto, Keisuke Imoto, Tatsuya Komatsu et al.

Synthesizing and converting environmental sounds have the potential for many applications such as supporting movie and game production, data augmentation for sound event detection and scene classification. Conventional works on synthesizing and converting environmental sounds are based on a physical modeling or concatenative approach. However, there are a limited number of works that have addressed environmental sound synthesis and conversion with statistical generative models; thus, this research area is not yet well organized. In this paper, we review problem definitions, applications, and evaluation methods of environmental sound synthesis and conversion. We then report on environmental sound synthesis using sound event labels, in which we focus on the current performance of statistical environmental sound synthesis and investigate how we should conduct subjective experiments on environmental sound synthesis.

Yoshiki Masuyama, Masahito Togami, Tatsuya Komatsu

In this paper, we propose two mask-based beamforming methods using a deep neural network (DNN) trained by multichannel loss functions. Beamforming technique using time-frequency (TF)-masks estimated by a DNN have been applied to many applications where TF-masks are used for estimating spatial covariance matrices. To train a DNN for mask-based beamforming, loss functions designed for monaural speech enhancement/separation have been employed. Although such a training criterion is simple, it does not directly correspond to the performance of mask-based beamforming. To overcome this problem, we use multichannel loss functions which evaluate the estimated spatial covariance matrices based on the multichannel Itakura--Saito divergence. DNNs trained by the multichannel loss functions can be applied to construct several beamformers. Experimental results confirmed their effectiveness and robustness to microphone configurations.

Chaitanya Narisetty, Tatsuya Komatsu, Reishi Kondo

This paper proposes a determined blind source separation method using Bayesian non-parametric modelling of sources. Conventionally source signals are separated from a given set of mixture signals by modelling them using non-negative matrix factorization (NMF). However in NMF, a latent variable signifying model complexity must be appropriately specified to avoid over-fitting or under-fitting. As real-world sources can be of varying and unknown complexities, we propose a Bayesian non-parametric framework which is invariant to such latent variables. We show that our proposed method adapts to different source complexities, while conventional methods require parameter tuning for optimal separation.

Chaitanya Narisetty, Tatsuya Komatsu, Reishi Kondo

This paper proposes an effective modelling of sound event spectra with a hidden data-size-imbalance, for improved Acoustic Event Detection (AED). The proposed method models each event as an aggregated representation of a few latent factors, while conventional approaches try to find acoustic elements directly from the event spectra. In the method, all the latent factors across all events are assigned comparable importance and complexity to overcome the hidden imbalance of data-sizes in event spectra. To extract latent factors in each event, the proposed method employs clustering and performs non-negative matrix factorization to each latent factor, and learns its acoustic elements as a sub-dictionary. Separate sub-dictionary learning effectively models the acoustic elements with limited data-sizes and avoids over-fitting due to hidden imbalances in training data. For the task of polyphonic sound event detection from DCASE 2013 challenge, an AED based on the proposed modelling achieves a detection F-measure of 46.5%, a significant improvement of more than 19% as compared to the existing state-of-the-art methods.