Research

Diffusion-based speech enhancement (SE) models need to incorporate correct prior knowledge as reliable conditions to generate accurate predictions. However, providing reliable conditions using noisy features is challenging. One solution is to use features enhanced by deterministic methods as conditions. However, the information distortion and loss caused by deterministic methods might affect the diffusion process. In this paper, we first investigate the effects of using different deterministic SE models as conditions for diffusion. We validate two conditions depending on whether the noisy feature was used as part of the condition: one using only the deterministic feature (deterministic-only), and the other using both deterministic and noisy features (deterministic-noisy). Preliminary investigation found that using deterministic enhanced conditions improves hearing experiences on real data, while the choice between using deterministic-only or deterministic-noisy conditions depends on the deterministic models. Based on these findings, we propose a dual-streaming encoding Repair-Diffusion Model for SE (DERDM-SE) to more effectively utilize both conditions. Moreover, we found that fine-grained deterministic models have greater potential in objective evaluation metrics, while UNet-based deterministic models provide more stable diffusion performance. Therefore, in the DERDM-SE, we propose a deterministic model that combines coarse- and fine-grained processing. Experimental results on CHiME4 show that the proposed models effectively leverage deterministic models to achieve better SE evaluation scores, along with more stable performance compared to other diffusion-based SE models.

While waveform-domain speech enhancement (SE) has been extensively investigated in recent years and achieves state-of-the-art performance in many datasets, spectrogram-based SE tends to show robust and stable enhancement behavior. In this paper, we propose a waveform-spectrogram hybrid method (WaveSpecEnc) to improve the robustness of waveform-domain SE. WaveSpecEnc refines the corresponding temporal feature map by spectrogram encoding in each encoder layer. Incorporating spectral information provides robust human hearing experience performance. However, it has a minor automatic speech recognition (ASR) improvement. Thus, we improve it for robust ASR by further utilizing spectrogram encoding information (WaveSpecEnc+) to both the SE front-end and ASR back-end. Experimental results using the CHiME-4 dataset show that ASR performance in real evaluation sets is consistently improved with the proposed method, which outperformed others, including DEMUCS and Conv-Tasnet. Refining in the shallow encoder layers is very effective, and the effect is confirmed even with a strong ASR baseline using WavLM.

Self-supervised learning (SSL)-based pretrained models have significantly improved automatic speech recognition (ASR) performance. As the feature extraction (FE) module has also been well-trained with a large amount of training data, freezing the FE during finetuning for downstream ASR tasks is common. When there is a severe mismatch between the simulated noisy data for pretraining and real noisy data, however, finetuning the FE with the real noisy data should be done without losing the effective information of the pretrained FE. In this paper, we propose a dual-path adaptation of the FE to address this problem. It combines the frozen pretrained FE path and the finetuned-adapted FE path with convolutional fusion layers. Moreover, adapters are inserted into the Transformer encoder. The experimental results using the CHiME–4 dataset show that using adapters for the FE or the Transformer encoder is effective, but achieving synergy of these two is challenging. Finetuning of the FE combined with adapters in the encoder realizes effective model adaptation. Moreover, the proposed method utilizes the complementarity between the pretrained and the finetuned FE paths, achieving significant improvements even with noise-robust WavLM models.

Diffusion-based generative speech enhancement (SE) has recently received attention, but reverse diffusion remains time-consuming. One solution is to initialize the reverse diffusion process with enhanced features estimated by a predictive SE system. However, the pipeline structure currently does not consider for a combined use of generative and predictive decoders. The predictive decoder allows us to use the further complementarity between predictive and diffusion-based generative SE. In this paper, we propose a unified system that use jointly generative and predictive decoders across two levels. The encoder encodes both generative and predictive information at the shared encoding level. At the decoded feature level, we fuse the two decoded features by generative and predictive decoders. Specifically, the two SE modules are fused in the initial and final diffusion steps: the initial fusion initializes the diffusion process with the predictive SE to improve convergence, and the final fusion combines the two complementary SE outputs to enhance SE performance. Experiments conducted on the Voice-Bank dataset demonstrate that incorporating predictive information leads to faster decoding and higher PESQ scores compared with other score-based diffusion SE (StoRM and SGMSE+).

Large language models (LLMs) provide strong linguistic priors for serialized output training (SOT), yet LLM-based multi-talker ASR degrades substantially as the number of overlapping talkers increases. Conventional systems expose acoustic evidence primarily through an initial projected mixture prefix, requiring the decoder to preserve and recover talker-relevant information indirectly throughout autoregressive generation. We first examine whether this limitation can be resolved by enriching the static prefix using discrete connectionist temporal classification (CTC) tokens, hybrid token--acoustic prompts, and continuous talker-specific representations. The results suggest that acoustic content alone does not fully address the conditioning bottleneck. We therefore extend onset-based serialization from the output target to the acoustic-conditioning pathway and introduce persistent decoder-side access to SOT-aligned serialized acoustic memory. Talker-specific representations are organized in utterance-onset order and retained as external acoustic memory, while the conventional mixture prefix provides a complementary global-conditioning path. LLM layers query this memory throughout generation through gated residual cross-attention. We further introduce a second adaptation stage that jointly applies low-rank updates to the acoustic-retrieval pathway and selected LLM self-attention projections. Experiments on LibriMix show consistent improvements over static-prefix prompting. These results indicate that effective LLM-based multi-talker ASR depends not only on providing richer acoustic representations, but also on maintaining persistent access to acoustic evidence structured according to the serialized output.

Large language models (LLMs) provide strong semantic priors that can improve multi-talker automatic speech recognition (MT-ASR), but using an LLM as an autoregressive decoder is computationally expensive and remains fragile under heavy overlap. In this paper, we propose an encoder-only MT-ASR framework that adapts an LLM to multi-talker conditioning and distills its semantic guidance into the encoder during training, while retaining fast CTC-style decoding at inference. Our model employs a post-encoder separator with serialized CTC to produce talker-ordered transcripts, and leverages an adapted LLM-based SOT objective as a multi-talker-aware teacher signal to explicitly regularize mixed-speech representations. To further support variable numbers of talkers, we introduce a Talker-Count Head that predicts the talker count and dynamically selects the appropriate decoding branch. Experiments on LibriMix show that the proposed encoder-only model achieves comparable performance to LLM-based systems in the two-talker condition, while delivering significant improvements in the three-talker condition with significant small RTF.

Prompts are crucial for task definition and for improving the performance of large language models (LLM)-based systems. However, existing LLM-based multi-talker (MT) automatic speech recognition (ASR) systems either omit prompts or rely on simple task-definition prompts, with no prior work exploring the design of prompts to enhance performance. In this paper, we propose extracting serialized output prompts (SOP) and explicitly guiding the LLM using structured prompts to improve system performance (SOP-MT-ASR). A Separator and serialized Connectionist Temporal Classification (CTC) layers are inserted after the speech encoder to separate and extract MT content from the mixed speech encoding in a first-speaking-first-out manner. Subsequently, the SOP, which serves as a prompt for LLMs, is obtained by decoding the serialized CTC outputs using greedy search. To train the model effectively, we design a three-stage training strategy, consisting of serialized output training (SOT) fine-tuning, serialized speech information extraction, and SOP-based adaptation. Experimental results on the LibriMix dataset show that, although the LLM-based SOT model performs well in the two-talker scenario, it fails to fully leverage LLMs under more complex conditions, such as the three-talker scenario. The proposed SOP approach significantly improved performance under both two- and three-talker conditions.

Serialized output training (SOT) attracts increasing attention due to its convenience and flexibility for multi-speaker automatic speech recognition (ASR). However, it is not easy to train with attention loss only. In this paper, we propose the overlapped encoding separation (EncSep) to fully utilize the benefits of the connectionist temporal classification (CTC) and attention (CTC-Attention) hybrid loss. This additional separator is inserted after the encoder to extract the multi-speaker information with CTC losses. Furthermore, we propose the serialized speech information guidance SOT (GEncSep) to further utilize the separated encodings. The separated streams are concatenated to provide single-speaker information to guide attention during decoding. The experimental results on Libri2Mix and Libri3Mix show that the single-speaker encoding can be separated from the overlapped encoding. The CTC loss helps to improve the encoder representation under complex scenarios (three-speaker and noisy conditions), which makes the EncSep have a relative improvement of more than 8% and 6% on the noisy Libri2Mix and Libri3Mix evaluation sets, respectively. GEncSep further improved performance, which was more than 12% and 9% relative improvement for the noisy Libri2Mix and Libri3Mix evaluation sets.

Conventional Japanese automatic speech recognition (ASR) is supervised by an orthographic transcript, although the same written form can correspond to different lexical readings realized in speech. Such utterances receive an identical target, so their reading distinction is absent from the supervision interface and cannot be recovered reliably by post-hoc text-only grapheme-to-phoneme conversion. We present Ruby-ASR, which refines the conventional target into a span-bound orthographic--lexical-reading sequence. Unlike separate full-sentence orthographic and phonological outputs, the ruby representation locally binds each written span to its realized reading and permits deterministic recovery of both views. We instantiate the target under subtitle-style and verbatim-style transcription conventions using a Qwen3-ASR backbone; a mora-level CTC objective provides auxiliary monotonic reading supervision. The experimental results across five Japanese benchmarks show that refining the recognition target can improve lexical-reading recovery without sacrificing readable orthographic transcription. We release the checkpoints and inference code.

While large language model (LLM)-based text-to-speech (TTS) systems have achieved high-quality speech synthesis, most existing systems focus on English and Chinese. Japanese, however, remains under-explored, and its unique linguistic challenges, such as widespread context-dependent kanji polyphony, have yet to be adequately tackled. Here we introduce Sarashina2.2-TTS (https://github.com/sbintuitions/sarashina2.2-tts), a Japanese-centric LLM-TTS system that tackles these challenges through a dual approach: data strategy and evaluation methodology. First, we scale training to approximately 361k hours of speech, incorporating a balanced mix of Japanese and English data. Furthermore, we design a targeted data augmentation pipeline covering all 2,136 Joyo (regular-use) kanji designated by Japan's Agency for Cultural Affairs to efficiently address kanji polyphony disambiguation. Second, we introduce the Joyo Kanji Yomi Benchmark (https://github.com/sbintuitions/JoyoKanji-Yomi-Benchmark), covering all 2,136 Joyo kanji and their 4,378 readings. Alongside this benchmark, we propose Kana-CER, a metric that compares synthesized speech against reference readings in the kana space, eliminating orthographic variations to directly measure pronunciation correctness. Experiments demonstrate that our targeted data augmentation significantly improves reading accuracy. Overall, Sarashina2.2-TTS achieves state-of-the-art kanji-level reading accuracy and matches top baselines on general sentence-level pronunciation, while delivering the highest speaker similarity in zero-shot Japanese speech synthesis. Furthermore, cross-lingual evaluation reveals that Sarashina2.2-TTS is the only system that maintains stable Japanese pronunciation regardless of the prompt language, confirming that our balanced training approach improves cross-lingual robustness.

SpeechLLMs typically combine ASR-trained encoders with text-based LLM backbones, leading them to inherit written-style output patterns unsuitable for text-to-speech synthesis. This mismatch is particularly pronounced in Japanese, where spoken and written registers differ substantially in politeness markers, sentence-final particles, and syntactic complexity. We propose a preference-based alignment approach to adapt Japanese SpeechLLMs for speech-worthy outputs: text that is concise, conversational, and readily synthesized as natural speech. To rigorously evaluate this task, we introduce SpokenElyza, a benchmark for Japanese speech-worthiness derived from ELYZA-tasks-100 with auditory verification by native experts. Experiments show that our approach achieves substantial improvement on SpokenElyza while largely preserving performance on the original written-style evaluation. We will release SpokenElyza to support future research on Japanese spoken dialog systems.

Dialogue systems based on large language models (LLMs) have advanced significantly in recent years. However, dialectal variation remains a major challenge, particularly for systems that process spoken input. LLM-based speech language models (SLMs), which integrate LLMs with speech processing components, show promise for spoken language tasks, yet their ability to comprehend dialects has not been sufficiently studied. Moreover, it remains unclear how the dialectal understanding of the base LLM affects SLM performance. This study investigates the dialectal robustness of both LLMs and SLMs using Japanese dialects as a test case. We define robustness as the ratio of performance on dialectal versus standard inputs, enabling fair comparisons. Our experiments show that SLM robustness correlates with that of their text-based counterparts. Furthermore, training with dialectal data and fine-tuning the speech encoder each improves robustness in SLMs.