Takafumi Moriya

Atsushi Ando, Ryo Masumura, Akihiko Takashima et al.

This paper investigates the effectiveness and implementation of modality-specific large-scale pre-trained encoders for multimodal sentiment analysis~(MSA). Although the effectiveness of pre-trained encoders in various fields has been reported, conventional MSA methods employ them for only linguistic modality, and their application has not been investigated. This paper compares the features yielded by large-scale pre-trained encoders with conventional heuristic features. One each of the largest pre-trained encoders publicly available for each modality are used; CLIP-ViT, WavLM, and BERT for visual, acoustic, and linguistic modalities, respectively. Experiments on two datasets reveal that methods with domain-specific pre-trained encoders attain better performance than those with conventional features in both unimodal and multimodal scenarios. We also find it better to use the outputs of the intermediate layers of the encoders than those of the output layer. The codes are available at https://github.com/ando-hub/MSA_Pretrain.

Kohei Matsuura, Takanori Ashihara, Takafumi Moriya et al.

End-to-end speech summarization (E2E SSum) is a technique to directly generate summary sentences from speech. Compared with the cascade approach, which combines automatic speech recognition (ASR) and text summarization models, the E2E approach is more promising because it mitigates ASR errors, incorporates nonverbal information, and simplifies the overall system. However, since collecting a large amount of paired data (i.e., speech and summary) is difficult, the training data is usually insufficient to train a robust E2E SSum system. In this paper, we present two novel methods that leverage a large amount of external text summarization data for E2E SSum training. The first technique is to utilize a text-to-speech (TTS) system to generate synthesized speech, which is used for E2E SSum training with the text summary. The second is a TTS-free method that directly inputs phoneme sequence instead of synthesized speech to the E2E SSum model. Experiments show that our proposed TTS- and phoneme-based methods improve several metrics on the How2 dataset. In particular, our best system outperforms a previous state-of-the-art one by a large margin (i.e., METEOR score improvements of more than 6 points). To the best of our knowledge, this is the first work to use external language resources for E2E SSum. Moreover, we report a detailed analysis of the How2 dataset to confirm the validity of our proposed E2E SSum system.

Takanori Ashihara, Takafumi Moriya, Kohei Matsuura et al.

Self-supervised learning (SSL) for speech representation has been successfully applied in various downstream tasks, such as speech and speaker recognition. More recently, speech SSL models have also been shown to be beneficial in advancing spoken language understanding tasks, implying that the SSL models have the potential to learn not only acoustic but also linguistic information. In this paper, we aim to clarify if speech SSL techniques can well capture linguistic knowledge. For this purpose, we introduce SpeechGLUE, a speech version of the General Language Understanding Evaluation (GLUE) benchmark. Since GLUE comprises a variety of natural language understanding tasks, SpeechGLUE can elucidate the degree of linguistic ability of speech SSL models. Experiments demonstrate that speech SSL models, although inferior to text-based SSL models, perform better than baselines, suggesting that they can acquire a certain amount of general linguistic knowledge from just unlabeled speech data.

Kohei Matsuura, Takanori Ashihara, Takafumi Moriya et al.

End-to-end speech summarization (E2E SSum) directly summarizes input speech into easy-to-read short sentences with a single model. This approach is promising because it, in contrast to the conventional cascade approach, can utilize full acoustical information and mitigate to the propagation of transcription errors. However, due to the high cost of collecting speech-summary pairs, an E2E SSum model tends to suffer from training data scarcity and output unnatural sentences. To overcome this drawback, we propose for the first time to integrate a pre-trained language model (LM), which is highly capable of generating natural sentences, into the E2E SSum decoder via transfer learning. In addition, to reduce the gap between the independently pre-trained encoder and decoder, we also propose to transfer the baseline E2E SSum encoder instead of the commonly used automatic speech recognition encoder. Experimental results show that the proposed model outperforms baseline and data augmented models.

Takafumi Moriya, Shota Horiguchi, Marc Delcroix et al.

Extending the RNN Transducer (RNNT) to recognize multi-talker speech is essential for wider automatic speech recognition (ASR) applications. Multi-talker RNNT (MT-RNNT) aims to achieve recognition without relying on costly front-end source separation. MT-RNNT is conventionally implemented using architectures with multiple encoders or decoders, or by serializing all speakers' transcriptions into a single output stream. The first approach is computationally expensive, particularly due to the need for multiple encoder processing. In contrast, the second approach involves a complex label generation process, requiring accurate timestamps of all words spoken by all speakers in the mixture, obtained from an external ASR system. In this paper, we propose a novel alignment-free training scheme for the MT-RNNT (MT-RNNT-AFT) that adopts the standard RNNT architecture. The target labels are created by appending a prompt token corresponding to each speaker at the beginning of the transcription, reflecting the order of each speaker's appearance in the mixtures. Thus, MT-RNNT-AFT can be trained without relying on accurate alignments, and it can recognize all speakers' speech with just one round of encoder processing. Experiments show that MT-RNNT-AFT achieves performance comparable to that of the state-of-the-art alternatives, while greatly simplifying the training process.

Kohei Matsuura, Takanori Ashihara, Takafumi Moriya et al.

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.

Atsunori Ogawa, Takafumi Moriya, Naoyuki Kamo et al.

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.

Takanori Ashihara, Takafumi Moriya, Kohei Matsuura et al.

Self-supervised learning (SSL) is seen as a very promising approach with high performance for several speech downstream tasks. Since the parameters of SSL models are generally so large that training and inference require a lot of memory and computational cost, it is desirable to produce compact SSL models without a significant performance degradation by applying compression methods such as knowledge distillation (KD). Although the KD approach is able to shrink the depth and/or width of SSL model structures, there has been little research on how varying the depth and width impacts the internal representation of the small-footprint model. This paper provides an empirical study that addresses the question. We investigate the performance on SUPERB while varying the structure and KD methods so as to keep the number of parameters constant; this allows us to analyze the contribution of the representation introduced by varying the model architecture. Experiments demonstrate that a certain depth is essential for solving content-oriented tasks (e.g. automatic speech recognition) accurately, whereas a certain width is necessary for achieving high performance on several speaker-oriented tasks (e.g. speaker identification). Based on these observations, we identify, for SUPERB, a more compressed model with better performance than previous studies.

Kenichi Fujita, Takanori Ashihara, Hiroki Kanagawa et al.

This paper proposes a zero-shot text-to-speech (TTS) conditioned by a self-supervised speech-representation model acquired through self-supervised learning (SSL). Conventional methods with embedding vectors from x-vector or global style tokens still have a gap in reproducing the speaker characteristics of unseen speakers. A novel point of the proposed method is the direct use of the SSL model to obtain embedding vectors from speech representations trained with a large amount of data. We also introduce the separate conditioning of acoustic features and a phoneme duration predictor to obtain the disentangled embeddings between rhythm-based speaker characteristics and acoustic-feature-based ones. The disentangled embeddings will enable us to achieve better reproduction performance for unseen speakers and rhythm transfer conditioned by different speeches. Objective and subjective evaluations showed that the proposed method can synthesize speech with improved similarity and achieve speech-rhythm transfer.

Ryo Masumura, Naoki Makishima, Taiga Yamane et al.

This paper proposes a novel automatic speech recognition (ASR) system that can transcribe individual speaker's speech while identifying whether they are target or non-target speakers from multi-talker overlapped speech. Target-speaker ASR systems are a promising way to only transcribe a target speaker's speech by enrolling the target speaker's information. However, in conversational ASR applications, transcribing both the target speaker's speech and non-target speakers' ones is often required to understand interactive information. To naturally consider both target and non-target speakers in a single ASR model, our idea is to extend autoregressive modeling-based multi-talker ASR systems to utilize the enrollment speech of the target speaker. Our proposed ASR is performed by recursively generating both textual tokens and tokens that represent target or non-target speakers. Our experiments demonstrate the effectiveness of our proposed method.

Takafumi Moriya, Takanori Ashihara, Masato Mimura et al.

A hybrid autoregressive transducer (HAT) is a variant of neural transducer that models blank and non-blank posterior distributions separately. In this paper, we propose a novel internal acoustic model (IAM) training strategy to enhance HAT-based speech recognition. IAM consists of encoder and joint networks, which are fully shared and jointly trained with HAT. This joint training not only enhances the HAT training efficiency but also encourages IAM and HAT to emit blanks synchronously which skips the more expensive non-blank computation, resulting in more effective blank thresholding for faster decoding. Experiments demonstrate that the relative error reductions of the HAT with IAM compared to the vanilla HAT are statistically significant. Moreover, we introduce dual blank thresholding, which combines both HAT- and IAM-blank thresholding and a compatible decoding algorithm. This results in a 42-75% decoding speed-up with no major performance degradation.

Takanori Ashihara, Marc Delcroix, Takafumi Moriya et al.

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.

Kenichi Fujita, Hiroshi Sato, Takanori Ashihara et al.

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.

Takanori Ashihara, Takafumi Moriya, Shota Horiguchi et al.

Target-speaker speech processing (TS) tasks, such as target-speaker automatic speech recognition (TS-ASR), target speech extraction (TSE), and personal voice activity detection (p-VAD), are important for extracting information about a desired speaker's speech even when it is corrupted by interfering speakers. While most studies have focused on training schemes or system architectures for each specific task, the auxiliary network for embedding target-speaker cues has not been investigated comprehensively in a unified cross-task evaluation. Therefore, this paper aims to address a fundamental question: what is the preferred speaker embedding for TS tasks? To this end, for the TS-ASR, TSE, and p-VAD tasks, we compare pre-trained speaker encoders (i.e., self-supervised or speaker recognition models) that compute speaker embeddings from pre-recorded enrollment speech of the target speaker with ideal speaker embeddings derived directly from the target speaker's identity in the form of a one-hot vector. To further understand the properties of ideal speaker embedding, we optimize it using a gradient-based approach to improve performance on the TS task. Our analysis reveals that speaker verification performance is somewhat unrelated to TS task performances, the one-hot vector outperforms enrollment-based ones, and the optimal embedding depends on the input mixture.

Atsunori Ogawa, Naoyuki Kamo, Kohei Matsuura et al.

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.

Atsushi Ando, Takafumi Moriya, Shota Horiguchi et al.

This paper presents a novel speaking-style captioning method that generates diverse descriptions while accurately predicting speaking-style information. Conventional learning criteria directly use original captions that contain not only speaking-style factor terms but also syntax words, which disturbs learning speaking-style information. To solve this problem, we introduce factor-conditioned captioning (FCC), which first outputs a phrase representing speaking-style factors (e.g., gender, pitch, etc.), and then generates a caption to ensure the model explicitly learns speaking-style factors. We also propose greedy-then-sampling (GtS) decoding, which first predicts speaking-style factors deterministically to guarantee semantic accuracy, and then generates a caption based on factor-conditioned sampling to ensure diversity. Experiments show that FCC outperforms the original caption-based training, and with GtS, it generates more diverse captions while keeping style prediction performance.

Takanori Ashihara, Takafumi Moriya, Kohei Matsuura et al.

Self-supervised learning (SSL) has been dramatically successful not only in monolingual but also in cross-lingual settings. However, since the two settings have been studied individually in general, there has been little research focusing on how effective a cross-lingual model is in comparison with a monolingual model. In this paper, we investigate this fundamental question empirically with Japanese automatic speech recognition (ASR) tasks. First, we begin by comparing the ASR performance of cross-lingual and monolingual models for two different language tasks while keeping the acoustic domain as identical as possible. Then, we examine how much unlabeled data collected in Japanese is needed to achieve performance comparable to a cross-lingual model pre-trained with tens of thousands of hours of English and/or multilingual data. Finally, we extensively investigate the effectiveness of SSL in Japanese and demonstrate state-of-the-art performance on multiple ASR tasks. Since there is no comprehensive SSL study for Japanese, we hope this study will guide Japanese SSL research.

Hiroshi Sato, Tsubasa Ochiai, Marc Delcroix et al.

The combination of a deep neural network (DNN) -based speech enhancement (SE) front-end and an automatic speech recognition (ASR) back-end is a widely used approach to implement overlapping speech recognition. However, the SE front-end generates processing artifacts that can degrade the ASR performance. We previously found that such performance degradation can occur even under fully overlapping conditions, depending on the signal-to-interference ratio (SIR) and signal-to-noise ratio (SNR). To mitigate the degradation, we introduced a rule-based method to switch the ASR input between the enhanced and observed signals, which showed promising results. However, the rule's optimality was unclear because it was heuristically designed and based only on SIR and SNR values. In this work, we propose a DNN-based switching method that directly estimates whether ASR will perform better on the enhanced or observed signals. We also introduce soft-switching that computes a weighted sum of the enhanced and observed signals for ASR input, with weights given by the switching model's output posteriors. The proposed learning-based switching showed performance comparable to that of rule-based oracle switching. The soft-switching further improved the ASR performance and achieved a relative character error rate reduction of up to 23 % as compared with the conventional method.

Tomohiro Tanaka, Ryo Masumura, Mana Ihori et al.

We propose a cross-modal transformer-based neural correction models that refines the output of an automatic speech recognition (ASR) system so as to exclude ASR errors. Generally, neural correction models are composed of encoder-decoder networks, which can directly model sequence-to-sequence mapping problems. The most successful method is to use both input speech and its ASR output text as the input contexts for the encoder-decoder networks. However, the conventional method cannot take into account the relationships between these two different modal inputs because the input contexts are separately encoded for each modal. To effectively leverage the correlated information between the two different modal inputs, our proposed models encode two different contexts jointly on the basis of cross-modal self-attention using a transformer. We expect that cross-modal self-attention can effectively capture the relationships between two different modals for refining ASR hypotheses. We also introduce a shallow fusion technique to efficiently integrate the first-pass ASR model and our proposed neural correction model. Experiments on Japanese natural language ASR tasks demonstrated that our proposed models achieve better ASR performance than conventional neural correction models.

Hiroshi Sato, Tsubasa Ochiai, Marc Delcroix et al.

Although recent advances in deep learning technology improved automatic speech recognition (ASR), it remains difficult to recognize speech when it overlaps other people's voices. Speech separation or extraction is often used as a front-end to ASR to handle such overlapping speech. However, deep neural network-based speech enhancement can generate `processing artifacts' as a side effect of the enhancement, which degrades ASR performance. For example, it is well known that single-channel noise reduction for non-speech noise (non-overlapping speech) often does not improve ASR. Likewise, the processing artifacts may also be detrimental to ASR in some conditions when processing overlapping speech with a separation/extraction method, although it is usually believed that separation/extraction improves ASR. In order to answer the question `Do we always have to separate/extract speech from mixtures?', we analyze ASR performance on observed and enhanced speech at various noise and interference conditions, and show that speech enhancement degrades ASR under some conditions even for overlapping speech. Based on these findings, we propose a simple switching algorithm between observed and enhanced speech based on the estimated signal-to-interference ratio and signal-to-noise ratio. We demonstrated experimentally that such a simple switching mechanism can improve recognition performance when processing artifacts are detrimental to ASR.