Let IT Begin

ZipCodec: Simple and Pretrained-Model-Free Speech Tokenizer via Flow-MatchingSpeech tokenizer는 massive pre-trained model이나 complex multi-stage training pipeline에 의존함ZipCodecZipformer-based flow-matching decoder에 acoustic reconstruction을 offloadingEncoder Consistency Regularization loss를 도입하여 encoder의 semantic capacity를 향상논문 (INTERSPEECH 2026) : Paper Link1. IntroductionSpeech tokenizer는 continuo..

ETC-TTS: Emotion Trajectory Learning for Controllable Emotional Text-to-SpeechEmotional Text-to-Speech는 여전히 intensity modeling의 한계가 있음ETC-TTSEmotion intensity를 latent style space의 continuous trajectory로 modelingResidual Vector Quantizer-based style extractor를 통해 emotion-aware representation을 제공하고 flow matching objective를 통해 neutral/target emotion 간의 trajectory를 학습논문 (INTERSPEECH 2026) : Paper Li..

CFLOW-VC: An Unsupervised Cycle Training Strategy based on Normalizing flows for Voice ConversionVoice Conversion model은 non-parallel setting에서 train-test mismatch 문제가 존재함CFLOW-VCNormalizing flow를 기반으로 expressive style feature를 capture 하는 mel-style encoder를 도입Noisy environment에 대한 robustness를 향상하기 위해 data augmentation을 수행논문 (INTERSPEECH 2026) : Paper Link1. IntroductionVoice Conversion (VC)는 ori..

Spiking Vocos: An Energy-Efficient Neural VocoderSpiking Neural Network를 사용하면 energy efficiency를 향상할 수 있음Spiking VocosMultiply-Accumulate operation을 reduce 하는 Spiking ConvNeXt module을 구성하고 signal dynamics를 preserve 하는 amplitude shortcut path를 incorporateSelf-architectural distillation strategy를 활용하여 performance gap을 줄이고 temporal dimension에서의 information fusion을 위한 lightweight Temporal Shift Modu..

BridgeCodec: Mamba Enhanced Neural Audio Codec with Schrodinger Bridge at Low BitrateNeural audio codec은 rigidly co-trained encoder-decoder pair로 인해 interoperability의 한계가 있음BridgeCodecCross-codec translation을 Schrodinger bridge로 formulate 하여 disparate distribution 간의 robust mapping을 지원추가적으로 Mamba-enhanced U-Net architecture를 활용하여 long range temporal dependency를 capture논문 (INTERSPEECH 2026) : Pap..

MSR-HuBERT: Self-Supervised Pre-Training for Adaptation to Multiple Sampling Rates기존의 Self-Supervised Learning model은 single sampling rate에서만 동작함MSR-HuBERTMulti-sampling-rate adaptive downsampling CNN을 활용하여 여러 sampling rate의 raw waveform을 shared temporal resolution으로 mappingUnified mixed-rate pre-training, fine-tuning을 지원하여 frequency modeling을 개선논문 (INTERSPEECH 2026) : Paper Link1. IntroductionS..

TASTE-Streaming: Towards Streamable Text-Aligned Speech Tokenization and Embedding for Spoken Language ModelingText-speech joint spoken language modeling은 modality mismatch 문제와 non-causal decoder로 인한 streaming의 한계가 있음TASTE-STASTE를 기반으로 instant dual-modality encoding을 위해 CTC-based ASR module을 integrateUnit decoder를 활용해 on-the-fly decoding을 지원하고 joint training을 채택하여 latency를 절감논문 (INTERSPEECH 2026..

RobustSpeechFlow: Learning Robust Text-to-Speech Trajectories via Augmentation-based Contrastive Flow MatchingFlow-matching Text-to-Speech는 imperfect alignment로 인한 content fidelity의 한계가 존재함RobustSpeechFlowLength-preserving repeat와 skip latent augmentation을 활용한 contrastive flow matching을 통해 alignment robustness를 향상External aligner와 preference data에 대한 의존성을 제거하고 기존 pipeline에 readily integrate논문 (I..

MeanVoiceFlow2: Joint Optimization of Mean Flow and Content Encoder for Fast One-Step Zero-Shot Voice ConversionFlow matching을 활용하면 high-quality voice conversion이 가능함MeanVoiceFlow2MeanVoiceFlow를 기반으로 content encoder의 computational bottleneck을 완화추가적으로 disentanglement를 향상하기 위해 sample mixing과 Diffusion-GAN training을 incorporate논문 (INTERSPEECH 2026) : Paper Link1. IntroductionVoice Conversion (VC)는 ..