The Inference Report

Full-duplex speech models can listen and speak simultaneously, enabling natural interaction, but become increasingly difficult to control as the conversation history grows. When used as user simulators, this lack of control can cause them to deviate from prescribed scenarios and produce unreliable evaluation outcomes. We introduce SimIF-Bench (Simulator Instruction-Following Benchmark), which evaluates whether a conversational model stays within a prescribed scenario and completes multiple goals in the required order. The benchmark reveals that current open-source full-duplex models struggle to follow such constraints. We then introduce a Group Reward-Decoupled Normalization Policy Optimization (GDPO)-based training recipe that enables a full-duplex model to follow textual instructions during an ongoing conversation while maintaining its turn-taking ability. By connecting the resulting SteerablePlex to an asynchronous backend language model that monitors the conversation and provides instructions when needed, we build a more controllable full-duplex user simulator that follows multi-stage constraints more reliably than existing open-source models and GPT-Realtime.

Blind Source Separation(BSS) is a fundamental problem in signal processing, aiming to separate multiple source signals from their mixtures without prior knowledge of the sources or the mixing process. Traditional approaches, such as Independent Vector Analysis (IVA) exploits statistical independence of sources. Recently, diffusion-based approaches have emerged as a promising alternative by leveraging powerful generative priors. Among them, ArrayDPS formulates BSS problem as a posterior sampling problem, and utilizes a pretrained speech diffusion model to guide the recovery of clean source signals. A key factor behind its separation capability is the multi-channel consistency (MC) objective, which enforces the estimated source signals to reconstruct the observed microphone mixtures through the estimated acoustic transfer functions. However, the number of microphones in the array is often limited, which constrains the performance of ArrayDPS. To address this issue, we propose VM-ArrayDPS, a novel method that augments the microphone array with virtual microphones with higher-SNR, these microphones can offer extra MC constraints to enhance the separation performance. Experimental results demonstrate that VM-ArrayDPS significantly outperforms ArrayDPS on both 2-speaker and 3-speaker datasets, showcasing the effectiveness of virtual microphone augmentation in improving BSS performance. We also did ablation studies to show the influence of the number of virtual microphones and weight of the MC objective brought by virtual microphones.

Contextual biasing supplies an ASR system with a list of expected words at inference time, but existing methods rely on word boundaries that Japanese and Chinese do not provide. We present a boundary-free biasing decoder for frozen public CTC models, built on a character-level Aho-Corasick automaton, with no training and no second pass. Two evidence-based mechanisms replace the boundary: a depth-adaptive gate that sets how hard to push from match depth, and reading-space matching for when the audio is right but the characters are wrong. On Aishell-1 NE's hard R1 subset we reach 66.5% recall, above the trained CLAS baseline (64%), transferring to WenetSpeech and to a second architecture without retuning. We release the first open Japanese contextual-biasing benchmark, where biasing lifts rare-word recall by 25 points at precision above 97%, and still by 19 and 22 points against 1,000-word lists.

Spoken language identification (LID) aims to recognize the target language regardless of accent. In practice, however, LID models fine-tuned from self-supervised speech representations frequently confuse accents with languages, misclassifying non-native (L2) speech as the speaker's first language (L1). We show that non-native speech representations lie between native target-language and native L1 poles, causing systematic misclassification. To address this, we introduce a geometric projection that estimates an L1-bias direction solely from native speech and removes it before the frozen LID head. Across five MMS-LID models and non-native corpora, this projection substantially improves target language identification for L2-accented speech while preserving predictions for native speech. These results show that accent-induced L1 bias can be corrected directly within the representation space without L2 training data or model adaptation.

When does distribution alignment help a frozen foundation-model embedding generalize across acoustic domains? For cross-domain mosquito-species classification we report a strength-monotonic law: the stronger an encoder is on the target task, the more its unseen-domain generalization relies on a distribution-alignment (MMD) term, and the more it is harmed by domain-rebalanced sampling. Across four encoder families and a within-encoder HuBERT layer sweep (n=8), the rebalancing leg orders exactly with encoder strength (Spearman -1.000), while the MMD-benefit leg is monotonic within each stream and -0.857 pooled; fixing architecture and varying only representation strength flips the rebalancing effect from benefit to collapse. The law is actionable: a single MMD term is the sole lever on a strong encoder, so we reduce the field's default recipe to a frozen Perch 2.0 embedding, a lightweight probe, cross-entropy, one MMD, and input augmentation. The reduced recipe stays within seed noise of the full composite (BA_unseen 0.299+/-0.006 vs. 0.307+/-0.014). As boundary conditions of the same law, three community defaults (backbone fine-tuning, multi-modal fusion, and domain rebalancing) each hurt unseen-domain accuracy under a leave-domain protocol, shown with single-variable, multi-seed evidence. We present a mechanism and the recipe it explains, not a leaderboard entry.

Audio benchmarks are built around short, pre-segmented clips, limiting model design to brief inputs or fixed vocabularies. To close this gap, we introduce Logbook, a benchmark for hour-scale audio understanding, with recordings ranging from ten minutes to six days. Given a continuous audio recording and an event label vocabulary, a system must predict a gap-free segmentation with an event label and a description per segment. We compare 52 systems, end-to-end and cascaded, and ablate fine-tuning, context length, and reasoning budget. We find the task tractable, though the best systems remain below the human reference. Also, over-segmentation is pervasive, and fine-tuning partially mitigates it. Finally, end-to-end are often better than cascaded systems, but degrades with longer context.

Prosodic stress is a crucial aspect of automatic pronunciation assessment (APA), encompassing both sentence stress detection (SSD) and word stress detection (WSD). SSD highlights semantically salient words that shape discourse meaning, while WSD identifies the primary stressed syllable within each word to ensure lexical clarity. However, most prior work treats SSD and WSD as independent tasks, overlooking their shared reliance on prosodic cues such as pitch, duration, and intensity. To address this gap, we propose an effective SSD approach combining SSD with auxiliary WSD via a novel modeling paradigm. In addition, we introduce a word-span stress regularizer (WSR) that concentrates token-level SSD probabilities within each stressed word span. Experiments on the TinyStress-15K benchmark show that the proposed method outperforms strong baselines, with the complete configuration achieving the best SSD result.

This paper presents the HINTT system submitted to the 2nd Challenge and Workshop on Multilingual Conversational Speech Language Model (MLC-SLM). We address multilingual speaker-attributed ASR, where systems must determine who spoke when and what was spoken. We investigate two modeling strategies for this problem: a cascaded pipeline that combines speaker diarization with speech-LLM-based ASR, and a unified speech LLM that directly generates speaker labels, timestamps, and transcriptions. Our final submission is based on the cascaded pipeline, consisting of a fine-tuned DiariZen diarization model, a fine-tuned Qwen3-ASR model, and LLM-based generative error correction. For comparison, we also fine-tune VibeVoice-ASR as a unified model using the same official training data. All task-specific fine-tuning and model selection are performed using only the official MLC-SLM data, without external data or pseudo-labels. Experimental results demonstrate that the cascaded system remains more reliable under the MLC-SLM Task 1 conditions, while unified speech LLMs offer a promising direction for future speaker-attributed ASR.

Full-duplex spoken dialogue models listen and speak at the same time, enabling voice agents to have natural, low-latency interactions that turn-based systems cannot offer. However, they are commonly evaluated against single-sided interlocutors: pre-recorded audio that cannot react, or an automated examiner that reacts in real time but only administers a fixed sequence of tests and is never graded. These single-sided frameworks evaluate only half of a two-body problem, where turn-taking, overlap, and interruption are joint products of two coupled speakers. We propose DyaFDB, a framework that evaluates full-duplex models in a dyadic setup: two models converse directly under assigned roles with cooperative or conflicting goals, and both sides are scored offline with an external judge. DyaFDB probes how the two models behave toward each other, such as how they take turns or carry an assigned role under different interests. We instantiate four tasks as 140 scenarios and record 7,560 conversations, covering six self- and cross-play pairings. Throughout the experiments, we observe that how a model behaves continually reshapes its partner. We thus demonstrate that each model must be both the examiner and examinee of the other, and no single fixed interlocutor can play both parts. We will release the scenarios, role prompts, and recording protocols between two full-duplex models, without any pre-recorded audio.

Code-switched (CS) speech leaks through the monolingual language identification (LID) filters used to curate massive speech corpora, calling for CS-aware LID (CS-LID). We formulate utterance-level CS-LID as multi-label language-set prediction and propose a set generator that directly outputs the languages in an utterance, comparing it against atomic-pair and score-based classification baselines. Oracle Top-k is the strongest baseline, but thresholding fails because no single threshold separates CS from monolingual speech. Our set generator predicts the correct language count on unseen pairs without assuming the number of languages, but underperforms oracle Top-k in exact set accuracy. Our analysis identifies the key obstacles to robust CS-LID: oracle cardinality, threshold instability, language bias in CS training data, and the synthetic-to-real gap.

Real-time voice conversion (VC) systems commonly rely on pretrained speaker embeddings from automatic speaker verification (ASV) models. While effective for speaker discrimination, these embeddings are trained to remain stable across phonetic and prosodic variations within-speaker, which may conflict with frame-level acoustic generation in streaming constraints. To address this issue, we propose VOSSA (Voiceprint Optimization for Streaming Speech Architectures), a speaker representation framework that extracts speaker information from intermediate content encoder layers and aggregates using attentive statistics pooling. The embedding is trained jointly with VC objectives, removing the need for a separate speaker encoder. Across six datasets, VOSSA improves F0 dynamics and vowel-discriminative acoustic cues while maintaining comparable NISQA-MOS, WER, and speaker similarity. Perceptual tests further indicate improvements in naturalness, speaker similarity, intelligibility, and vibrancy.

Explicit prosodic cues may help automatic speech recognition (ASR) of spontaneous speech, but auxiliary representations typically require additional trainable components, making it unclear whether gains come from the auxiliary information or the fusion mechanism. We address this using a frozen HuBERT backbone and a 64-dimensional representation trained to predict log F0, voicing, Delta log F0, log energy, and spectral tilt. We compare a frozen-backbone recognizer (Baseline), trainable fusion with zero auxiliary input (Null), and the same fusion supplied with the learned representation (Learned). Across Buckeye, Switchboard, and AMI IHM, Null reduces WER by 0.71-1.45 points over Baseline, whereas Learned differs from Null by +0.07, -0.09, and +0.00 points, with no significant differences. However, removing or mismatching the representation at inference increases Learned WER. Thus, Learned depends on the representation yet shows no measurable incremental WER benefit over the parameter-matched control.

Speech self-supervised learning aims to learn general-purpose representations for downstream speech tasks. However, current approaches rely on complex, carefully designed prediction targets. We challenge this necessity with GLaS-JEPA, a framework that directly predicts the current encoder's continuous representations at masked positions, without contrastive learning, discrete targets, or separate EMA target encoders. We prevent representation collapse using SIGReg representation-space regularization, eliminating the need for engineered target-generation mechanisms. Pretrained on 960 hours of LibriSpeech, our 57M-parameter model achieves a 6.89% WER on frozen-encoder SUPERB ASR and a 25.87% CER on slot filling, outperforming the best non-distilled sub-90M baselines by 43.1% and 22.0%, respectively. These results demonstrate that highly competitive speech representations can emerge from a radically simplified training recipe.

Contemporary deep speech enhancement (SE) models are often trained with specific auxiliary terms in the loss function as a way to improve their performance in terms of perceptual metrics. Nevertheless, a higher score on a perceptual metric does not necessarily correlate with an improved listening experience. Through objective and subjective experiments, we assess the performance of SE models trained with two different types of auxiliary PESQ loss terms. The numerical evaluation on a suite of standard metrics suggests that, while models optimized for PESQ naturally obtain higher PESQ scores in the test set, for most other metrics the scores do not significantly change. In some cases, the PESQ loss even results in worse PESQ scores on mismatched data. A formal listening experiment reveals that the models without a PESQ loss were generally preferred over models that include it, across all settings. Finally, we analyze the relative importance of PESQ in the composite metrics CSIG, CBAK and COVL, and find that PESQ dominates all of them. Our study highlights the perils of over-reliance on PESQ and stresses the importance of a complete evaluation procedure for SE.

We address Task A of the 1st DAFx Parameter Estimation Challenge, which aims to retrieve the physical parameters of a plate model from an impulse response. To do so, we use the Simulation-Based Inference (SBI) framework, in which we train a neural network to estimate a density over plate parameters given an impulse response, using a dataset generated by the simulator. Inference for a new impulse response then requires only a forward pass through the network, without involving the simulator. For each test observation, we fine-tune a specific network: additional simulation rounds are performed by sampling parameters from the current estimated distribution, simulating the corresponding impulse responses, and fine-tuning to produce the specialized network.

Speech-based screening is a promising, non-invasive approach for detecting Alzheimer's disease and related cognitive risks. However, models trained on a single domain often generalize poorly to unseen languages, tasks, or recording protocols. This paper investigates this deployment gap using a leave-one-corpus-out evaluation across four distinct datasets. Among 70 interpretable speech and language features, 59 exhibit direction conflicts between healthy control and cognitive risk groups across corpora, with pause, silence, and speech rate showing high protocol sensitivity. Furthermore, while the XLM-R text baseline achieves strong average performance, its Area Under the ROC Curve (AUC) drops to 0.520 on the weakest held-out domain. A standard GroupDRO baseline reaches a 0.766 mean speaker AUC and a 0.504 worst-domain AUC under the same protocol. To address this, we propose a fusion method that integrates XLM-R text baseline scores with evidence anchors selected during training. Balanced fusion achieves a 0.785 mean speaker AUC, while anchor-heavy fusion raises the worst-case speaker AUC to 0.615. This work highlights the need to audit feature transferability and report worst-case domain robustness in cognitive speech screening.

Self-supervised learning (SSL) countermeasures (CMs) have shown strong performance in recent years. However, they often show degraded performance while facing unseen spoofing attacks and mismatched conditions. This study examines the Voxtral audio-language model (ALM) framework for spoofing detection, as a step toward combining CM capabilities within the ALM framework. We analyze how Voxtral captures spoofing cues through audio-text processing and propose an instruction-guided approach that uses label-sequence likelihoods to evaluate bonafide and spoofed speech. Experiments on the ASVspoof databases show that without task-specific adaptation, the LLM layers emphasize semantic representations, reducing the separability of spoof-discriminative acoustic cues compared to the Whisper-based audio encoder. Consequently, spoofing-related information becomes less separable after language-model processing. We also applied lightweight adaptation using weight-decomposed low-rank adaptation (DoRA) to the Voxtral model and propose the Spooftral model, achieving an equal error rate (EER) of 4.25% on the ASVspoof5 evaluation set.

Full-duplex dialogue systems, which listen while speaking, must distinguish a completed turn from a pause within a turn and an interruption that requests a turn from a brief acknowledgment or speech addressed to a third party. Yet existing conversational corpora provide limited control over these events and limited labels for their intent. We present a pipeline for synthesizing intent-labeled, two-channel conversational speech from relational event lists. An LLM authors each event's speaker, text, conversational act, and attachment to an earlier event without predicting absolute timestamps. Events are synthesized independently, aligned with their source text, and placed on a shared clock, so turn-taking landmarks are measured from the rendered signal while silence durations are specified or sampled from turn-taking distributions. The pipeline covers 42 phenomena across eight families in English and Mandarin, derives frame-level system actions from authored intent, and promotes diversity using small, diverse sets of prior examples and batch prompts that request alternatives with self-reported probabilities. Ablations show gains in each targeted diversity dimension. On a four-action label space for taking, holding, releasing, and not holding the conversational floor, a semantic voice-activity detector using only current and past audio reaches start-speaking and start-listening F1 scores of 0.819 and 0.802. When generating its own responses, the full-duplex speech model Moshi takes 0.85 of the reference turns after fine-tuning on the generated corpus, compared with 0.44 before fine-tuning. Its frame-level precision for predicting system-floor occupancy rises from 0.46 to 0.88. With reference context at each step, its frame-level floor F1 rises from 0.893 to 0.962. These results show that controlled synthesis can provide learnable and transferable supervision for full-duplex turn management.

New words are invented every day. A human listener can learn a new word by hearing it clearly once and inferring its usage from sentence context. This paper proposes granting ASR a similar ability to learn the contextual representations and spellings of new words from unlabeled test data at test time. A frozen CTC acoustic model provides spellings, a frozen language model provides contextual evidence for out-of-vocabulary (OOV) word detection, and an adaptation module expands the vocabulary by learning the lexical token representations with distributions over CTC-generated candidates. The spelling model of each token is optimized by minimizing a Kullback-Leibler divergence (KLD) objective. We demonstrate that the CTC-weighted language model log likelihood ratio can be interpreted as the KLD between the unknown correct ASR and the unsupervised learned ASR, and that, using a Pinsker bound, the square root of KLD can be interpreted as an upper bound on the total variation distance between the true and estimated spelling of the unknown word. Experiments show relative OOV character-error-rate reductions of up to 14.97% on LibriSpeech and 6.67% on dysarthric Speech Accessibility Project data for recurring OOV words, relative to the corresponding rescoring system.

Post-training quantization (PTQ) reduces the cost of on-device text-to-speech (TTS), but published evaluations cover one system or method. We evaluate PTQ across TTS architectures under one protocol with three core models, weight and activation ablations of eight more, and two held-out models quantized blind. Four-bit per-channel weights reduce UTMOS, a predicted mean opinion score, by 2.8 on Supertonic and 0.07 on Kokoro, and per-tensor scaling can cause severe degradation even at 8 bits. The same bit width yields different outcomes, because the sensitive component is model-specific and not reliably predicted from the model class. A staged ablation procedure identifies it, and per-layer GPTQ can restore it to within 0.1 UTMOS. Real int8 and int4 kernels reproduce the simulated ordering at hardware-dependent cost. On a Mac mini, a 4-bit weight kernel runs Supertonic at 0.60x the fp32 latency while int8 is slower, so each configuration requires validation on the target runtime.

We present a personalized Korean visual speech recognition (VSR) system and quantify, on the nine-camera OLKAVS corpus, the gap between the population-level benchmark score and an individual user's error. A video-only Conformer initialized from English-trained weights attains 9.95 - 12.19% character error rate (CER) under the corpus protocol against the published 26.64, and 19.00 - 21.52 on unseen wording. Per speaker, CER spans 1.0 to 52.2%, with seen wording lowering CER by 7.0 - 9.0 points and professional delivery and spontaneous speech raising it by 8.5 - 10.5 and 12.7 points. A low-rank adapter with 4.6% of the parameters, trained on 4 to 29 minutes of the user's frontal video, lowers the CER of twelve high-error speakers by 2.13 to 3.58 points, transfers to every camera without loss, and keeps 85% of the full fine-tuning gain at 12% of its cost to other speakers. Cameras above the mouth plane add about six CER points as a constant offset that training on all views keeps small.

The acoustic front-end determines which forensic cues a speech deepfake detector can exploit. The wavelet scattering transform (WST) provides stable multiscale coefficients with explicit coordinates, yet direct flattening obscures the parent relation between paths. We introduce WST-Graph, reconstructing these paths as a sparse modulation-carrier grid for an AASIST graph backend. Modulation-level normalization and length-aware adaptive local attention pooling produce fixed relative-time representations while retaining the acoustic axes before learned adaptation. This yields a waveform-to-graph interface with a fixed, parameter-free WST. Our configurations remain competitive with AASIST while using approximately 60% fewer trainable parameters and show clear gains on selected out-of-domain benchmarks. These results underscore the value of preserving parent-child relations within the carrier-modulation topology when constructing a compact, physically grounded interface for graph-based speech deepfake detection. Code will be released at https://github.com/saki-ciallo/wst-graph.

We adapt Reinforce Adjoint Matching (RAM), a reward-based post-training method, to generative speech enhancement (SE). Starting from a pretrained SE model, RAM tilts the model's conditional distribution toward outputs with higher reward. During training, the current model generates enhanced speech on-policy, evaluates each generated endpoint with a potentially non-differentiable reward, and analytically re-noises the endpoint to construct inputs for a reward-guided regression objective. This enables post-training directly on real recordings using weak supervision, such as text transcripts, without requiring paired clean speech targets or reward gradients. We investigate word error rate (WER)-based post-training and whether recognition performance can be improved without compromising perceptual speech quality. Experiments on real CHiME-4 recordings reduce WER by 5.08 percentage points relative to pretrained FlowSE without reducing any of the reported non-intrusive speech quality metrics. A subjective listening test at the default reward scale finds no statistically significant preference between the post-trained and pretrained models.

Cross-speaker acoustic-to-articulatory inversion requires accounting for anatomical differences between speakers. We propose a geometric adaptation framework that uses anatomical landmarks, primarily on vertebrae and dental structures,to transfer predictions from a fixed inversion model to unseen speakers. An affine transformation followed by thin-plate spline (TPS) deformation maps the predicted contours of 10 vocal-tract structures into each target speaker's geometry without retraining. Landmarks are identified in one selected /u/ frame per speaker as a common phonetic reference without assuming identical articulatory configurations across speakers, and the resulting mapping is reused across recordings. We train the model on a single-speaker rt-MRI database and evaluate adaptation on eight speakers from a separate multi-speaker rt-MRI database. We compare affine and TPS configurations using 12 or 14 landmarks. Affine12+TPS14 achieves the lowest mean point-to-closest-point error of 3.19mm. These results support the combined value of anatomical landmark information and nonrigid alignment.

Deploying real-time speech enhancement on resource-constrained devices requires meeting strict latency, memory, and energy constraints. Microcontroller NPUs can accelerate neural inference under these constraints, but only through a restricted set of operators in static, integer-quantized graphs. Recent speech-enhancement networks have reduced parameter counts and MACs to levels nominally suitable for microcontrollers, but their operators and execution patterns often remain incompatible with restricted NPUs. We address this gap by redesigning LiSenNet, a 37k parameter sub-band dual-path model, for the STM32N6570-DK Neural-ART accelerator. We replace its recurrent bottleneck with convolutional frequency and temporal mixers, reformulate unsupported operations as static int8-compatible primitives, and use bounded decoder activations to preserve quality after quantization. On VoiceBank-DEMAND, the final NPU-compatible model matches or exceeds the recurrent LiSenNet baseline, reaching PESQ 3.08 versus 3.01 in FP32 and 3.01 versus 2.93 in int8. Deployed on a microcontroller, it processes each 16 ms input hop in 4.83 ms, corresponding to a real-time factor of 0.30. Stateless receptive-field recomputation is an order of magnitude slower at the same frame rate despite higher accelerator utilization. These results show that parameter count and operator compatibility, quantization range, and persistent streaming state must be co-designed to achieve efficient real-time speech enhancement on restricted NPUs.

Deep learning-based speech enhancement is increasingly deployed on-device in hearing aids, headsets, and earbuds. Most of these devices, however, can only accelerate static int8 graphs, so a depth-varying network must be implemented as several graphs, orchestrated by a policy. In this paper, we supervise every intermediate depth of one causal model, then we fine-tune its output heads to guarantee that deeper outputs are never worse than shallower ones. Using this training protocol, we can derive a family of static models that are more Pareto-efficient than their equivalently-sized counterparts trained from scratch on the same budget. Specifically, we achieve up to 0.11 higher PESQ for equivalent compute, and match the best PESQ at 30% less compute. We then quantize the models to int8 and measure the latency-quality frontier on an STM32N6 microcontroller. On VoiceBank-DEMAND, the dynamic enhancer lies on the same frontier as the static models, rather than trading quality for dynamic execution. Running the policy on the companion Cortex-M55 takes only 26 $μ$s per frame, while splitting the enhancer into separate NPU graphs adds 2.2% latency overhead. The cost of dynamic execution is therefore small.

Real-time voice assistants must reason over evolving requests, execute actions, and follow conversational rules. Qwen-Audio-3.1-Realtime brings these requirements together through Think, Act, and Speak and Coordinate. Think combines Core-Cocktail supervised fine-tuning with Multimodality and Multi-Teacher On-Policy Distillation (M$^{2}$-OPD) to transfer language capabilities and develop native audio skills. Act uses self-evolving executable environments and multi-granularity rollouts for Group Relative Policy Optimization (GRPO), teaching the model to use tools, interpret feedback, and complete tasks. Speak and Coordinate aligns how, when, and whether the assistant speaks or acts. We evaluate audio reasoning, multilingual understanding, tool use, conversational behavior, full-duplex interaction, and safety. Compared with Qwen-Audio-3.0-Realtime, 3.1 raises overall task success from 78.4% to 82.0% on our half-duplex speech-to-text adaptation of $τ$-Voice. On speech-to-speech Full-Duplex-Bench v1.5, the response rate to background speech falls from 73.0% to 13.0%. We also present a separate Voice Harness prototype, using Qwen-Audio-3.0-Realtime as its foreground, that extends spoken interaction to persistent tasks through foreground--background coordination and memory.

Classifier-free Guidance (CFG) is widely adopted in text-to-speech (TTS) systems to enhance generation quality and conditioning fidelity by interpolating between conditioned and unconditioned predictions. A common unconditional technique is to use an empty representation, in the form of a fixed null vector. In this work, we propose replacing this representation with a learnable unconditional embedding, optimized to represent a meaningful unconditional state. Objective and subjective evaluations demonstrate that learnable null embeddings consistently outperform fixed null embeddings across speaker similarity, speech stability, and expressiveness, while exhibiting greater robustness to larger guidance scales. We further show that learning a distinct unconditional embedding for each of the TTS conditioning modalities allows fine-grained control over speaker and text guidance, showcasing the trade-off between similarity and quality, and stability and expressiveness in the generated speech.

Speech enhancement (SE) models typically rely on supervised learning with paired data examples where clean speech is synthetically degraded. This paradigm limits performance in real-world scenarios where the target environment's specific acoustic characteristics are unknown. We propose a fully unpaired SE framework that uses principled Diffusion Schrödinger Bridges (DSB) to learn a stochastic transport process between a clean and a degraded speech distribution. Algorithms for learning transport maps are computationally heavy since they require simulating differential equations during training, usually at each training step. Therefore, we propose using a high-efficiency Mamba Diffusion Model designed for end-to-end waveform processing. We compare against state-of-the-art methods for speech enhancement, both paired and unpaired, as well as a classical signal processing algorithm. Experimental results show that we are on par or better than the baselines while being orders of magnitude faster during inference. Furthermore, we show that the flexibility of the DSB formulation allows our model to generalize across SE tasks, offering a robust and efficient solution for real-world speech restoration.

Automatic lip-synchronous dubbing requires a speech synthesis model to generate alternating voice and silence patterns in the target language that match the timing of the source clip precisely to ensure an optimal viewing experience. Prior works address this problem by conditioning the speech synthesis process on lip movements extracted from the video signal. In this work, we condition the speech generation on a binary voice-activity signal, which has a lightweight representation and can be produced in multiple ways. We show that the model follows the voice-activity signal with high accuracy while maintaining natural prosody and semantically appropriate pause placement within sentences, as demonstrated through extensive objective and subjective evaluations. By randomly masking this condition during training, we make the feature entirely optional during inference, allowing editors to enforce or relax lip-sync constraints when desired.