Self-supervised learning (SSL) models such as WavLM have brought substantial improvements to speaker diarization by providing rich contextual representations. However, the high computational and memory costs of these models hinder their deployment in real-time and resource-constrained scenarios. In this work, we present a comprehensive study on compressing SSL-based diarization models through structured pruning guided by knowledge distillation. Building upon our previous work, we extend the analysis to include pruning objectives based on multiply-accumulate operations (MACs), investigate module-wise and progressive pruning strategies, and examine the impact of training data quantity. Experimental results show that our method reduces model size by up to 80% without degrading performance, achieving up to 4x faster inference on a single GPU. We further perform large-scale evaluations on a diverse compound dataset comprising eight public diarization corpora, where our best pruned model achieves state-of-the-art performance across most conditions. Additionally, we show strong generalization to the CHiME-6 dataset, attaining performance comparable to the third-place system in the CHiME-7 challenge without any domain adaptation. All models and code are publicly released to support reproducibility and future research.
In this paper, we propose a novel neural speaker diarization system using memory-aware multi-speaker embedding with sequence-to-sequence architecture (NSD-MS2S), which integrates a memory-aware multi-speaker embedding module with a sequence-to-sequence architecture. The system leverages a memory module to enhance speaker embeddings and employs a Seq2Seq framework to efficiently map acoustic features to speaker labels. Additionally, we explore the application of mixture of experts in speaker diarization, and introduce a Shared and Soft Mixture of Experts (SS-MoE) module to further mitigate model bias and enhance performance. Incorporating SS-MoE leads to the extended model NSD-MS2S-SSMoE. Experiments on multiple complex acoustic datasets, including CHiME-6, DiPCo, Mixer 6 and DIHARD-III evaluation sets, demonstrate meaningful improvements in robustness and generalization. The proposed methods achieve state-of-the-art results, showcasing their effectiveness in challenging real-world scenarios.
This paper presents Seewo's systems for both tracks of the Multilingual Conversational Speech Language Model Challenge (MLC-SLM), addressing automatic speech recognition (ASR) and speaker diarization with ASR (SD-ASR). We introduce a multi-stage training pipeline that explicitly enhances reasoning and self-correction in speech language models for ASR. Our approach combines curriculum learning for progressive capability acquisition, Chain-of-Thought data augmentation to foster intermediate reflection, and Reinforcement Learning with Verifiable Rewards (RLVR) to further refine self-correction through reward-driven optimization. This approach achieves substantial improvements over the official challenge baselines. On the evaluation set, our best system attains a WER/CER of 11.57% for Track 1 and a tcpWER/tcpCER of 17.67% for Track 2. Comprehensive ablation studies demonstrate the effectiveness of each component under challenge constraints.
We propose Speaker-Conditioned Serialized Output Training (SC-SOT), an enhanced SOT-based training for E2E multi-talker ASR. We first probe how SOT handles overlapped speech, and we found the decoder performs implicit speaker separation. We hypothesize this implicit separation is often insufficient due to ambiguous acoustic cues in overlapping regions. To address this, SC-SOT explicitly conditions the decoder on speaker information, providing detailed information about "who spoke when". Specifically, we enhance the decoder by incorporating: (1) speaker embeddings, which allow the model to focus on the acoustic characteristics of the target speaker, and (2) speaker activity information, which guides the model to suppress non-target speakers. The speaker embeddings are derived from a jointly trained E2E speaker diarization model, mitigating the need for speaker enrollment. Experimental results demonstrate the effectiveness of our conditioning approach on overlapped speech.
In the field of speaker diarization, the development of technology is constrained by two problems: insufficient data resources and poor generalization ability of deep learning models. To address these two problems, firstly, we propose an automated method for constructing speaker diarization datasets, which generates more accurate pseudo-labels for massive data through the combination of audio and video. Relying on this method, we have released Multi-modal, Multi-scenario and Multi-language Speaker Diarization (M3SD) datasets. This dataset is derived from real network videos and is highly diverse. In addition, we further propose a scenario-related model fine-tuning strategy. Based on the general model pre-trained using the above dataset, we combine the specific data of the target scenario (e.g., meetings) and achieve targeted optimization by using Adapter and LoRA joint fine-tuning, thus achieving the model's domain adaptation. Our dataset and code have been open-sourced at https://huggingface.co/spaces/OldDragon/m3sd.
Multi-speaker automatic speech recognition (MS-ASR) faces significant challenges in transcribing overlapped speech, a task critical for applications like meeting transcription and conversational analysis. While serialized output training (SOT)-style methods serve as common solutions, they often discard absolute timing information, limiting their utility in time-sensitive scenarios. Leveraging recent advances in large language models (LLMs) for conversational audio processing, we propose a novel diarization-aware multi-speaker ASR system that integrates speaker diarization with LLM-based transcription. Our framework processes structured diarization inputs alongside frame-level speaker and semantic embeddings, enabling the LLM to generate segment-level transcriptions. Experiments demonstrate that the system achieves robust performance in multilingual dyadic conversations and excels in complex, high-overlap multi-speaker meeting scenarios. This work highlights the potential of LLMs as unified back-ends for joint speaker-aware segmentation and transcription.
Self-supervised learning (SSL) models like WavLM can be effectively utilized when building speaker diarization systems but are often large and slow, limiting their use in resource constrained scenarios. Previous studies have explored compression techniques, but usually for the price of degraded performance at high pruning ratios. In this work, we propose to compress SSL models through structured pruning by introducing knowledge distillation. Different from the existing works, we emphasize the importance of fine-tuning SSL models before pruning. Experiments on far-field single-channel AMI, AISHELL-4, and AliMeeting datasets show that our method can remove redundant parameters of WavLM Base+ and WavLM Large by up to 80% without any performance degradation. After pruning, the inference speeds on a single GPU for the Base+ and Large models are 4.0 and 2.6 times faster, respectively. Our source code is publicly available.
In recent years, end-to-end approaches have made notable progress in addressing the challenge of speaker diarization, which involves segmenting and identifying speakers in multi-talker recordings. One such approach, Encoder-Decoder Attractors (EDA), has been proposed to handle variable speaker counts as well as better guide the network during training. In this study, we extend the attractor paradigm by moving beyond direct speaker modeling and instead focus on representing more detailed `speaker attributes' through a multi-stage process of intermediate representations. Additionally, we enhance the architecture by replacing transformers with conformers, a convolution-augmented transformer, to model local dependencies. Experiments demonstrate improved diarization performance on the CALLHOME dataset.
In speaker diarization, traditional clustering-based methods remain widely used in real-world applications. However, these methods struggle with the complex distribution of speaker embeddings and overlapping speech segments. To address these limitations, we propose an Overlapping Community Detection method based on Graph Attention networks and the Label Propagation Algorithm (OCDGALP). The proposed framework comprises two key components: (1) a graph attention network that refines speaker embeddings and node connections by aggregating information from neighboring nodes, and (2) a label propagation algorithm that assigns multiple community labels to each node, enabling simultaneous clustering and overlapping community detection. Experimental results show that the proposed method significantly reduces the Diarization Error Rate (DER), achieving a state-of-the-art 15.94% DER on the DIHARD-III dataset without oracle Voice Activity Detection (VAD), and an impressive 11.07% with oracle VAD.
End-to-end speaker diarization enables accurate overlap-aware diarization by jointly estimating multiple speakers' speech activities in parallel. This approach is data-hungry, requiring a large amount of labeled conversational data, which cannot be fully obtained from real datasets alone. To address this issue, large-scale simulated data is often used for pretraining, but it requires enormous storage and I/O capacity, and simulating data that closely resembles real conversations remains challenging. In this paper, we propose pretraining a model to identify multiple speakers from an input fully overlapped mixture as an alternative to pretraining a diarization model. This method eliminates the need to prepare a large-scale simulated dataset while leveraging large-scale speaker recognition datasets for training. Through comprehensive experiments, we demonstrate that the proposed method enables a highly accurate yet lightweight local diarization model without simulated conversational data.
We introduce VoxRAG, a modular speech-to-speech retrieval-augmented generation system that bypasses transcription to retrieve semantically relevant audio segments directly from spoken queries. VoxRAG employs silence-aware segmentation, speaker diarization, CLAP audio embeddings, and FAISS retrieval using L2-normalized cosine similarity. We construct a 50-query test set recorded as spoken input by a native English speaker. Retrieval quality was evaluated using LLM-as-a-judge annotations. For very relevant segments, cosine similarity achieved a Recall@10 of 0.34. For somewhat relevant segments, Recall@10 rose to 0.60 and nDCG@10 to 0.27, highlighting strong topical alignment. Answer quality was judged on a 0--2 scale across relevance, accuracy, completeness, and precision, with mean scores of 0.84, 0.58, 0.56, and 0.46 respectively. While precision and retrieval quality remain key limitations, VoxRAG shows that transcription-free speech-to-speech retrieval is feasible in RAG systems.
Human speech contains paralinguistic cues that reflect a speaker's physiological and neurological state, potentially enabling non-invasive detection of various medical phenotypes. We introduce the Human Phenotype Project Voice corpus (HPP-Voice): a dataset of 7,188 recordings in which Hebrew-speaking adults count for 30 seconds, with each speaker linked to up to 15 potentially voice-related phenotypes spanning respiratory, sleep, mental health, metabolic, immune, and neurological conditions. We present a systematic comparison of 14 modern speech embedding models, where modern speech embeddings from these 30-second counting tasks outperform MFCCs and demographics for downstream health condition classifications. We found that embedding learned from a speaker identification model can predict objectively measured moderate to severe sleep apnea in males with an AUC of 0.64 $\pm$ 0.03, while MFCC and demographic features led to AUCs of 0.56 $\pm$ 0.02 and 0.57 $\pm$ 0.02, respectively. Additionally, our results reveal gender-specific patterns in model effectiveness across different medical domains. For males, speaker identification and diarization models consistently outperformed speech foundation models for respiratory conditions (e.g., asthma: 0.61 $\pm$ 0.03 vs. 0.56 $\pm$ 0.02) and sleep-related conditions (insomnia: 0.65 $\pm$ 0.04 vs. 0.59 $\pm$ 0.05). For females, speaker diarization models performed best for smoking status (0.61 $\pm$ 0.02 vs 0.55 $\pm$ 0.02), while Hebrew-specific models performed best (0.59 $\pm$ 0.02 vs. 0.58 $\pm$ 0.02) in classifying anxiety compared to speech foundation models. Our findings provide evidence that a simple counting task can support large-scale, multi-phenotypic voice screening and highlight which embedding families generalize best to specific conditions, insights that can guide future vocal biomarker research and clinical deployment.
This paper describes the speaker diarization system developed for the Multimodal Information-Based Speech Processing (MISP) 2025 Challenge. First, we utilize the Sequence-to-Sequence Neural Diarization (S2SND) framework to generate initial predictions using single-channel audio. Then, we extend the original S2SND framework to create a new version, Multi-Channel Sequence-to-Sequence Neural Diarization (MC-S2SND), which refines the initial results using multi-channel audio. The final system achieves a diarization error rate (DER) of 8.09% on the evaluation set of the competition database, ranking first place in the speaker diarization task of the MISP 2025 Challenge.
Meetings are a valuable yet challenging scenario for speech applications due to complex acoustic conditions. This paper summarizes the outcomes of the MISP 2025 Challenge, hosted at Interspeech 2025, which focuses on multi-modal, multi-device meeting transcription by incorporating video modality alongside audio. The tasks include Audio-Visual Speaker Diarization (AVSD), Audio-Visual Speech Recognition (AVSR), and Audio-Visual Diarization and Recognition (AVDR). We present the challenge's objectives, tasks, dataset, baseline systems, and solutions proposed by participants. The best-performing systems achieved significant improvements over the baseline: the top AVSD model achieved a Diarization Error Rate (DER) of 8.09%, improving by 7.43%; the top AVSR system achieved a Character Error Rate (CER) of 9.48%, improving by 10.62%; and the best AVDR system achieved a concatenated minimum-permutation Character Error Rate (cpCER) of 11.56%, improving by 72.49%.
Speaker diarization, the process of identifying "who spoke when" in audio recordings, is essential for understanding classroom dynamics. However, classroom settings present distinct challenges, including poor recording quality, high levels of background noise, overlapping speech, and the difficulty of accurately capturing children's voices. This study investigates the effectiveness of multi-stage diarization models using Nvidia's NeMo diarization pipeline. We assess the impact of denoising on diarization accuracy and compare various voice activity detection (VAD) models, including self-supervised transformer-based frame-wise VAD models. We also explore a hybrid VAD approach that integrates Automatic Speech Recognition (ASR) word-level timestamps with frame-level VAD predictions. We conduct experiments using two datasets from English speaking classrooms to separate teacher vs. student speech and to separate all speakers. Our results show that denoising significantly improves the Diarization Error Rate (DER) by reducing the rate of missed speech. Additionally, training on both denoised and noisy datasets leads to substantial performance gains in noisy conditions. The hybrid VAD model leads to further improvements in speech detection, achieving a DER as low as 17% in teacher-student experiments and 45% in all-speaker experiments. However, we also identified trade-offs between voice activity detection and speaker confusion. Overall, our study highlights the effectiveness of multi-stage diarization models and integrating ASR-based information for enhancing speaker diarization in noisy classroom environments.
Speaker diarization is a fundamental task in speech processing that involves dividing an audio stream by speaker. Although state-of-the-art models have advanced performance in high-resource languages, low-resource languages such as Kurdish pose unique challenges due to limited annotated data, multiple dialects and frequent code-switching. In this study, we address these issues by training the Wav2Vec 2.0 self-supervised learning model on a dedicated Kurdish corpus. By leveraging transfer learning, we adapted multilingual representations learned from other languages to capture the phonetic and acoustic characteristics of Kurdish speech. Relative to a baseline method, our approach reduced the diarization error rate by seven point two percent and improved cluster purity by thirteen percent. These findings demonstrate that enhancements to existing models can significantly improve diarization performance for under-resourced languages. Our work has practical implications for developing transcription services for Kurdish-language media and for speaker segmentation in multilingual call centers, teleconferencing and video-conferencing systems. The results establish a foundation for building effective diarization systems in other understudied languages, contributing to greater equity in speech technology.
While voice technologies increasingly serve aging populations, current systems exhibit significant performance gaps due to inadequate training data capturing elderly-specific vocal characteristics like presbyphonia and dialectal variations. The limited data available on super-aged individuals in existing elderly speech datasets, coupled with overly simple recording styles and annotation dimensions, exacerbates this issue. To address the critical scarcity of speech data from individuals aged 75 and above, we introduce SeniorTalk, a carefully annotated Chinese spoken dialogue dataset. This dataset contains 55.53 hours of speech from 101 natural conversations involving 202 participants, ensuring a strategic balance across gender, region, and age. Through detailed annotation across multiple dimensions, it can support a wide range of speech tasks. We perform extensive experiments on speaker verification, speaker diarization, speech recognition, and speech editing tasks, offering crucial insights for the development of speech technologies targeting this age group.
In this paper, we introduce a multi-talker distant automatic speech recognition (DASR) system we designed for the DASR task 1 of the CHiME-8 challenge. Our system performs speaker counting, diarization, and ASR. It handles various recording conditions, from diner parties to professional meetings and from two to eight speakers. We perform diarization first, followed by speech enhancement, and then ASR as the challenge baseline. However, we introduced several key refinements. First, we derived a powerful speaker diarization relying on end-to-end speaker diarization with vector clustering (EEND-VC), multi-channel speaker counting using enhanced embeddings from EEND-VC, and target-speaker voice activity detection (TS-VAD). For speech enhancement, we introduced a novel microphone selection rule to better select the most relevant microphones among the distributed microphones and investigated improvements to beamforming. Finally, for ASR, we developed several models exploiting Whisper and WavLM speech foundation models. We present the results we submitted to the challenge and updated results we obtained afterward. Our strongest system achieves a 63% relative macro tcpWER improvement over the baseline and outperforms the challenge best results on the NOTSOFAR-1 meeting evaluation data among geometry-independent systems.
Speech technologies are transforming interactions across various sectors, from healthcare to call centers and robots, yet their performance on African-accented conversations remains underexplored. We introduce Afrispeech-Dialog, a benchmark dataset of 50 simulated medical and non-medical African-accented English conversations, designed to evaluate automatic speech recognition (ASR) and related technologies. We assess state-of-the-art (SOTA) speaker diarization and ASR systems on long-form, accented speech, comparing their performance with native accents and discover a 10%+ performance degradation. Additionally, we explore medical conversation summarization capabilities of large language models (LLMs) to demonstrate the impact of ASR errors on downstream medical summaries, providing insights into the challenges and opportunities for speech technologies in the Global South. Our work highlights the need for more inclusive datasets to advance conversational AI in low-resource settings.
The aim of this paper is to investigate the benefit of combining both language and acoustic modelling for speaker diarization. Although conventional systems only use acoustic features, in some scenarios linguistic data contain high discriminative speaker information, even more reliable than the acoustic ones. In this study we analyze how an appropriate fusion of both kind of features is able to obtain good results in these cases. The proposed system is based on an iterative algorithm where a LSTM network is used as a speaker classifier. The network is fed with character-level word embeddings and a GMM based acoustic score created with the output labels from previous iterations. The presented algorithm has been evaluated in a Call-Center database, which is composed of telephone interview audios. The combination of acoustic features and linguistic content shows a 84.29% improvement in terms of a word-level DER as compared to a HMM/VB baseline system. The results of this study confirms that linguistic content can be efficiently used for some speaker recognition tasks.