Papers

Speech-pretrained models have shown promise in animal bioacoustics, but the factors governing their cross-species transfer remain poorly understood. We evaluate 15 frozen encoders spanning monolingual and multilingual speech, speaker verification, animal bioacoustics, and general audio on caller identification across four species and call-type classification across three. Differences between the strongest speech- and animal-pretrained representations range from $-0.027$ to $+0.119$ UAR; speech is significantly better in three of seven settings ($p<0.01$) and not significantly different in the remaining four. Controlled comparisons show no systematic advantage from increased language coverage or animal-domain pretraining, while frozen speaker-verification embeddings transfer poorly. The results also show that transfer is strongly layer-dependent: raw-waveform models peak early, whereas patch-spectrogram models peak deeper.

Compact time-frequency separators that mask the mixture and refine through a shared cell face two limits. First, a bounded multiplicative mask only scales a mixture bin, so where overlapping components cancel, the estimate stays small. Second, a shared cell applies the same weights to every time-frequency token at every step, so enlarging it adds compute everywhere. We present SEAL (Sparse Expert routing with Additive Latent reconstruction) to address both. For reconstruction, a zero-sum additive residual bounded by the local mixture amplitude lets estimates be nonzero where components cancel yet still sum to the mixture. For routing, a query built from acoustic and inter-step evidence sends each token to one of six residual experts, and a norm cap keeps the step cue from overriding clear acoustic evidence. On EchoSet, SEAL (small) surpasses TIGER (small) by 0.31 dB SI-SDRi with 28% fewer parameters and 2.9 times fewer MACs, and SEAL (large) is within 0.07 dB SI-SDRi of TIGER (large) at 3.1 times fewer MACs.

Visual speech promises noise-robust keyword spotting, yet a visual stream is not necessarily used. On a tri-modal query-by-example keyword spotting (QbyE-KWS) benchmark, we find that a system with a task-trained visual encoder comes within 2 percentage points of a text-and-audio system in equal error rate (EER) at -10 dB, and link this gap to the encoder's lack of phonemic information. We present GIVE-KWS, whose fusion stage, GIVE (Gated Injection of Visual Evidence), conditions query audio on lip motion through gated cross-attention. We show that visual robustness depends on two interacting conditions: a phoneme-bearing visual representation, and fusion that injects visual evidence rather than rescaling audio features. Under a phoneme-bearing encoder, injection yields an effective SNR gain of 4.0-9.3 dB over masking at -10 dB, whereas under a phoneme-poor one it nearly vanishes. Relative to the benchmark system, GIVE-KWS reduces unseen-keyword EER by 72.9% at -10 dB and 62.8% on average.

We distill Kokoro-82M, a widely used open text-to-speech model with 54 voices, into Paradee, an 8.07M-parameter model that speaks one of them. Paradee keeps Kokoro's architecture with much narrower layers, and each of its two halves is trained separately against the frozen teacher. It has 10x fewer parameters and needs 15x less compute. We first synthesize a corpus with the teacher and keep its durations, pitch, energy and phoneme features. We then train a small text side to predict these values, and a small decoder to turn the teacher's saved values into the teacher's audio, first with spectral losses and then adversarially. Finally, we connect the two halves and quantize the weights to int8. It needs no alignment learning and no joint training, and it runs on one laptop. Stored in int8, Paradee is 8.5 MB, runs 25x faster than real time on one CPU thread, and scores 4.41 on UTMOS against the teacher's 4.52. The student initially kept a slight buzz, which we trace to the phase of voiced speech between 2 and 8 kHz. A phase-locking filter applied after synthesis removes most of it, with no training and no extra parameters. Code, model files and audio samples are at https://github.com/sahilmahendrakar/paradee

Objective audio quality metrics typically provide point estimates, whereas listening tests yield score distributions from which mean opinion scores (MOS), confidence intervals (CIs), and significance decisions are derived. We propose a lightweight intrusive metric that combines a PEAQ-style perceptual front-end (ITU-R BS.1387) with a bagging ensemble of regressors. Calibration with subjective data aligns ensemble outputs with listener scores. The resulting item-dependent score distributions enable uncertainty assessment, panel-size-matched CIs, and identification of less conclusive predictions. Calibration improves agreement with subjective distributions and CI coverage across all evaluated datasets while preserving MOS accuracy. Using only 11 fixed PEAQ features, low-capacity regressors, and public training data, the method performs comparably to more data-intensive end-to-end approaches. Its output can support uncertainty-aware assessment and target listening tests towards uncertain conditions. The distributions also enable approximate pairwise comparisons, but not yet reliable significance inference.

Sound morphing, generating intermediate sounds that transition from one sonic identity to another, can be a powerful tool for musical sound design. Existing diffusion-based morphing approaches entangle temporal structure and timbral identity, offering no mechanism to hold one fixed while transforming the other. We present smorph, a training-free guidance framework that preserves how a sound behaves over time while transforming what the sound is, allowing users to morph, for instance from brass to strings at a fixed pitch. We demonstrate across three morphing modes: prompt-to-prompt, audio-to-prompt, and audio-to-audio. Evaluations across diverse datasets show that smorph effectively produces smooth morph trajectories while substantially improving temporal-structure and source preservation over baselines, albeit with more conservative target-ward transformation in some settings. In an exploratory case study, musicians found smorph trajectories to be expressive and playable, suggesting structural anchoring can serve as a productive constraint for instrumental interaction.

Pretrained speech representations are increasingly used for depression detection, but selecting the best encoder layer on the same data used for evaluation biases reported performance. On DAIC-WOZ, with repeated nested cross-validation across five deep encoder families, naive best-of-25-layer probing inflates AUC by up to $0.09$. On shuffled labels over the real latent representations, naive best-of-25 still reaches AUC $0.59$ versus $0.50$ under nested selection, and this bias grows with the number of probed layers and with smaller samples, isolating selection, not signal, as the cause. The effect replicates on a second clinical corpus and language (Androids, Italian). Under leakage-free selection, a fair comparison across representation families identifies a compact, task-aligned affect–prosody model as competitive with substantially more complex SSL, deep-learning, audio-LLM, and codec-based approaches, while using $\sim\!10^{3}\times$ fewer trainable downstream parameters. Extending the analysis to individual PHQ-8 symptoms shows that the same selection bias persists at this finer-grained level. We argue nested model-selection protocols should be standard when probing encoder layer representations on small clinical speech datasets.

Low-power speech recognition requires reducing audio acquisition and feature-processing costs as well as neural inference. We present Lyric, a speech-recognition front end that uses wave-domain computing to extract features before digitization. Our proposed system uses passive acoustic resonators to separate speech by frequency and supplies their slowly varying envelopes to a compact temporal neural network. This moves spectral filtering into the acoustic structure, reducing the data and processing needed for recognition. We build a prototype and evaluate spoken-digit recognition on a speakerdisjoint AudioMNIST split across nine classifier families. The front end acquires 32 times fewer scalar samples than a 16-kHz waveform. In the EdgeSpeechNet-A comparison on a Raspberry Pi 4, we reduce preprocessing-plus-inference latency and estimated processor energy per inference by 98.6%, with a 3.58-percentage-point decrease in accuracy to 95.70%.

Voice anonymization (VA) is used to conceal the voice identities in speech data. Performance is usually estimated using automatic speaker verification (ASV) as a proxy to judge the capability of an attacker to re-identify speakers after anonymization. The defender anonymizes test utterances; the attacker compares them to equally anonymized reference utterances to infer voice identity, with degraded ASV performance indicating successful anonymization. Thus, the defender's protection is one-sided and only applied to test utterances via VA. Though unexplored so far, there is an opportunity to enhance anonymization by protecting reference utterances too. We present a new, preemptive VA paradigm: reference utterances are protected using adversarial noise to degrade their potential to infer voice identity. Preemptive protection does not degrade the perceived quality of reference utterances and is independent of the specific ASV system used for identity inference. By combining preemptive protection and regular test-side VA, ASV equal error rates can be increased from 14% to 45% (near-perfect privacy).

A dubbed character keeps its identity while the voice actor changes, so character identity and speaker identity are distinct properties of one recording, yet speaker verification measures only the latter. To address this gap, we propose KyaraBench, a benchmark that scores character voice directly instead of speaker voice, built from 85 human-audited identities in a dubbed anime corpus. It poses two challenging conditions: one that swaps the performer under a fixed character, and one that fixes the performer under changing characters. Listening studies with 78 participants provide human reference scores for cross-performer verification and same-actor discrimination. Speaker-verification baselines show increased errors under these character-specific conditions. We then train KyaraEmbed, a compact encoder using multilingual character supervision, same-actor negatives, and a language-alignment term. The model achieves the best performance on all character-specific conditions in the main comparison. We release the benchmark, protocol, and encoder publicly.

Dysarthric speech reconstruction (DSR) typically relies on linguistic or phonetic representations extracted from impaired speech to generate a more intelligible waveform. We investigate a complementary approach that instead intervenes in a representation related to speech production. Starting from a neural analysis–synthesis framework, we introduce a residual mapper that modifies an articulatory-aligned latent space while preserving speaker and prosodic information. The mapper is pretrained on parallel synthetic healthy and artificially dysarthric speech, then adapted to natural dysarthric speech using phoneme-guided and adversarial objectives. We further compare the articulatory bottleneck with a dimension-matched unsupervised representation to assess the benefit of explicit articulatory supervision.

We ask: given a retrieved source audio $S$ and a separate reference audio $R$, can we synthesize novel audio $Y$ out of this pair $(S,R)$ such that $Y$ remains acoustically consistent with $S$, while not persistently copying segments of $S$ or $R$? The first clause is a well-known goal in Foley audio production, and the second is a well-known issue in neural RAG when $S$ and $R$ are naively injected into neural generators. We show that both clauses can be addressed simultaneously using a method we coin relational synthesis, a variation of concatenative synthesis where target cost is replaced by a relational Gromov-like structural cost. Rather than imitating the content of $R$, relational synthesis exploits it from the "other side of the hill": it transfers the temporal structure and directed amplitude motion of $R$ to reorganize and concatenate the grains of $S$ in a novel manner that protects $S$'s acoustic information. Our experiments show that relational synthesis integrates naturally with neural RAG and produces Foley audio that performs well on metrics measuring temporal agreement, acoustic fidelity, and leakage persistence, while maintaining distribution-level quality and text alignment.

This paper revisits frame-wise saliency for audio moment retrieval (AMR). We show that the frame-wise saliency sequence, conventionally used only as an auxiliary output in DETR-based AMR models, can itself serve as an effective source of moment predictions. We convert the saliency sequence into ranked moments using a simple SED-inspired segmentation rule with no learned parameters, enabling moment retrieval directly from frame-wise temporal information. On the CASTELLA dataset, saliency-based prediction consistently outperforms decoder-based prediction from the same model across all 18 runs of QD-DETR and CG-DETR. For QD-DETR, simply replacing the inference output improves R1@0.7 from 21.0 to 36.1. The advantage remains 7-15 points when the two outputs are evaluated at their independently selected best epochs. The same tendency extends to TaskWeave and UVCOM, whereas TR-DETR shows the opposite behavior, suggesting that how saliency construction may matter. The performance gap is especially pronounced for short moments: for queries whose annotated moments average at most 2 s, R1@0.7 improves from 6.3 to 27.8 with QD-DETR. Decoder supervision nevertheless benefits saliency-based prediction, indicating that its role during training differs from the utility of its inference output.

Speaker tracking is the task of separating and following multiple speakers over time. It must address overlapped speech, speech onset and offset, talker identity, and time-varying speaker count. We propose a segment-online, modular framework for single- and multi-channel speaker tracking. The proposed system first performs speaker separation in each segment and computes speech activity via voice activity detection (VAD). To generate speaker tracks over time, we introduce a two-stage sequential organization strategy: Overlap-based stitching for continuous grouping and memory-based speaker verification for discontinuous grouping. For speaker separation, we employ complex spectral mapping to estimate the real and imaginary spectrograms of underlying speakers. The proposed system achieves state-of-the-art segment-online tracking performance on the LibriCSS and AMI datasets. Our framework significantly reduces diarization error rate (DER) and concatenated minimum-permutation word error rate (cpWER) compared to other methods.

We formulate joint estimation of common-slope decay rates and amplitudes from room impulse responses (RIRs) as parameterized nonnegative matrix factorization with the Itakura–Saito divergence as the loss function (IS-NMF). Estimation at each short-time Fourier transform frequency bin produces detailed reverberation time (RT) curves directly from RIR powers with no backward integration needed. Standard space-alternating generalized expectation-maximization (SAGE) algorithm yields closed-form amplitude updates and a convex subproblem for each decay rate update. To accelerate estimation, we introduce contribution-weighted SAGE, which emphasizes observations where each component contributes strongly to the modeled power. Experiments with synthetic data show accurate recovery of well-separated decays and faster loss reduction than standard SAGE. Application to measured coupled-room RIRs yields frequency-dependent RT curves and reveals complementary space-time contributions of the shared decay components.

Frozen speech foundation models (SFMs) make downstream adaptation efficient: the backbone can stay fixed while a task learns layer fusion and a lightweight classifier. Full adversarial fine-tuning is a standard route to robustness, but generating adversarial examples and updating the backbone for every task sacrifices that efficiency. We ask whether robustness can instead be learned before future tasks are known. For a frozen backbone and linear classifier, robustness can be understood through the interaction between representation stability and decision-boundary margin. This leads directly to our design: we stabilize representations across the hidden layers, rather than only the final layer, while preserving clean representations; after clean adaptation selects the layer mixture, we keep it fixed and enlarge only the classifier margin, without downstream adversarial examples. We evaluate Wav2Vec2, HuBERT, and WavLM Large on four tasks under adaptive 30 dB attacks. Across 12 backbone-task pairs, hierarchical robustification improves robust accuracy by 46.4 pp, while margin refinement adds 4.0 pp for 1.1 pp of clean accuracy. Code and configurations are available at https://github.com/arefmousavi/hierarchical-robust-sfm.

We propose UltraM2M, a weakly-supervised speech enhancement algorithm building upon the recent unsupervised mixture-to-mixture (M2M) algorithm. M2M realizes unsupervised speech enhancement by training deep neural networks on a set of real-recorded noisy-reverberant multi-channel mixture signals to estimate target speech and non-target signals. The two estimated signals are penalized by a so-called mixture-constraint (MC) loss, which constrains them to reconstruct the observed mixture signals. Although shown to be effective, the mixture constraint may be too weak to enable sufficient noise reduction. To deal with this, UltraM2M extends M2M by further leveraging text transcripts of real-recorded mixtures to design an automatic speech recognition (ASR) loss to penalize the estimated speech signal. The ASR loss can be viewed as a form of weak supervision that could help unsupervised enhancement. Evaluation results on the CHiME-4 dataset show the effectiveness of UltraM2M.

Learning-based acoustic sound source localization and detection (SSLD) requires large labeled datasets covering diverse acoustic conditions. Since obtaining measured data is costly, training commonly uses simulated data, while practical devices must operate under real-world conditions. However, higher simulation complexity increases data-generation cost, and the complexity required for reliable generalization remains unclear. In practice, SSLD methods often rely on efficient geometrical acoustic simulation, typically the image-source method. This work investigates how geometrical acoustic simulation complexity affects the real-world performance of a common multi-source SSLD model. We train the model with simulators ranging from anechoic conditions to high-order image-source simulations, optionally including diffuse reverberation, array simulation, and randomized image-source positions, and evaluate them on three measurement-based test datasets. Results show that anechoic simulation is insufficient, while medium-complexity image-source simulations already provide strong real-world performance. Further increases in complexity yield only marginal gains, with the best performance obtained using the highest tested image-source order, array simulation, and randomized image-source positions. For this configuration, measured-domain performance approaches within-domain simulated performance, suggesting limited benefit from further increasing simulation complexity. These findings show how the complexity-performance trade-off can be exploited in future data-driven SSLD, and highlight image-source randomization as an efficient way to improve generalization.

Musical-instrument recognition is a fundamental music information retrieval (MIR) task supporting transcription, source separation, and audio understanding. Despite significant algorithmic progress, there is still room for improvement in recognizing acoustically similar instruments, particularly in polyphonic mixtures. We examine the contribution of ultrasonic components, i.e., frequencies above the audible 0-20 kHz range. Although most of this content cannot be represented at conventional 44.1/48-kSPS sampling rates under the Nyquist criterion, it is available in high-quality 96-kSPS recordings. Using a 96-kSPS multitrack corpus covering 15 classes across performers, instruments, studios, and sessions, we compare full-band, low-pass, and ultrasonic-only inputs using classical machine-learning and deep-learning classifiers. Our results show that source-dependent ultrasonic extensions and broadband transients contain instrument-specific information that significantly improves isolated and polyphonic recognition and motivates wider-bandwidth studies of other MIR tasks.

Self-supervised learning (SSL) is standard for speech representation learning, but mainstream models are designed around single-speaker audio, limiting their usefulness in multi-speakers scenarios. We present SepRQ, an open-source SSL framework that replaces masked prediction with a pseudo-source-separation objective over frozen random-projection codebooks. By adopting a novel mask-free, multiresolution approach, SepRQ achieves state-of-the-art performance in Speaker Diarization and Speech Separation on the SUPERB benchmark, surpassing WavLM and other cocktail-party derived SSLs at both Base and Large scales, while requiring only 85.68M inference parameters. SepRQ also demonstrates strong performance across target-speaker tasks requiring enrollment (such as Target-Speaker Automatic Speech Recognition), and on the challenging multi-domain DIHARD 3 diarization dataset. Notably, we report strong separation capabilities on three-speaker mixtures (WSJ0-3Mix), where current SSL literature struggles. While cocktail-party SSLs remain scarce and closed-source, limited to C-HuBERT and the enrollment-based SA-WavLM, we open-source SepRQ to the community.

Activation steering has proven effective for controlling the behavior of Large Language Models (LLMs) at inference time, but its application to SpeechLLMs remains new, and training-free steering approaches for such models are still largely unexplored. We propose a training-free Contrastive Activation Addition (CAA) protocol that derives steering vectors for common speech tasks (e.g. transcription) in SpeechLLMs from a small number of labeled utterances. We showcase that adding these vectors in the representation space, at inference time, enforces better the targeted speech task. We further show that, when combined with prompting, these vectors yield to consistent improvement over prompting alone on most evaluated tasks such as Automatic Speech Recognition (ASR) or Emotion Recognition (ER), and transfer to out-of-domain data. We additionally demonstrate the usefulness of script-normalization directions to enforce the target script of a specific language.

Neural audio codecs compress waveforms into compact discrete tokens that underpin speech language models, real-time communication, and large-scale audio storage. Almost every dominant design, including residual vector quantization, finite scalar quantization, and single-codebook variants, follows the VQ-VAE template by partitioning the encoder latent through a learned codebook or a fixed scalar grid. We ask whether this partition is necessary. We introduce GS-Codec, a neural speech codec whose bottleneck is a parametric signal decomposition rather than a quantizer. We adapt Gaussian splatting from 3D scene reconstruction to one-dimensional latents. An inner optimization loop fits each encoder segment as a weighted sum of 1D Gaussian primitives. The decoder then reconstructs the waveform from the rendered sum. To avoid the cost of this iterative inner loop at inference time, we additionally train a lightweight GS Predictor Net that regresses the primitive parameters in a single forward pass. The encoder and decoder are trained end-to-end through the inner loop, with no quantizer anywhere in the training pipeline: the bottleneck is the decomposition itself, and scalar quantization is applied only post-training to the fitted parameters. Rather than relying on discrete codebook stages for bitrate control, our representation exposes a fine-grained rate-quality tradeoff: a single trained checkpoint supports post-training bitrate control by varying the number of primitives and the per-parameter bit depth, with no retraining required. GS-Codec matches or exceeds well-established open-source codecs such as EnCodec and DAC on speaker similarity (SIM), intelligibility (STOI), and perceptual quality (UTMOS) at comparable bitrates, while achieving comparable semantic performance (WER). Code and audio samples are available at https://ronaluf.github.io/gs-codec/

Although reinforcement learning (RL) post-training repairs the localized segmental errors of zero-shot text-to-speech (TTS), arriving at a working recipe still relies on tedious manual tuning, and whether LLM agents can take over this research pipeline is unclear. We investigate this question with AgenticTTS-Forge, a collaborative workflow that structures human guidance and agentic execution around a shared workspace, applied to CosyVoice2-0.5B. To measure what the agent automates, we audit its trajectory stage by stage against the published recipe. To measure what it exploits, we score its policies with held-out observers hidden from the agent. Our results show that the agent recovers an underspecified recipe, improves it, and, when gains stall, surveys the literature unprompted and pivots from the LM carrier to the flow carrier, halving Bad cases. However, its autonomy exposes three traps across the data, proxy, and algorithm axes: the held-out set leaks through a channel the contract never reads, a self-shaped reward inflates the proxy where it is scored, and separately tuned policies do not compose additively. These findings show that the binding constraint is measurement rather than reasoning, and can inform the design of harnesses whose contracts read every channel the agent does.

Delayed Streams Modeling (DSM) enables LLM-based streaming automatic speech recognition (ASR) by aligning acoustic and text streams on a common timeline. DSM adds the padding token <p> and the word-start token <w> to the LLM vocabulary and predicts them together with normal text tokens using the same softmax. We first show that <w> can be removed while maintaining competitive recognition performance. Based on this result, we propose Factorized DSM (F-DSM), which separates the waiting probability for <p> from the distribution over the original LLM vocabulary. This factorization removes ASR-specific tokens from the text prediction space and allows the large-vocabulary softmax to be skipped on waiting steps. Experiments on the Corpus of Spontaneous Japanese and LibriSpeech show that F-DSM achieves better recognition performance than DSM. It also greatly reduces GPU memory use while maintaining similar training throughput, provides a small inference speed improvement through softmax skipping, and reduces the degradation in text-only perplexity observed with DSM.

Auditory attention decoding (AAD) is often evaluated on static, simplified speech scenes that poorly match everyday listening. We introduce MOV-AAD, a large-scale dataset for studying auditory attention under moving, naturalistic conversations. MOV-AAD combines 64-channel EEG with synchronized physiological recordings, including eye tracking, respiration, galvanic skin response, heart rate, peripheral oxygen saturation, body temperature, body motion, and photoplethysmography, enabling analysis of cross-modal neural and physiological markers of attention and listening effort. This dataset uses a more ecologically valid listening paradigm with dynamically moving conversational speech sources and behavioral measures of attentional engagement. MOV-AAD supports research on robust AAD in realistic spatial dynamics, multimodal attention modeling, listening effort, and intersubject neural responses, providing a resource for benchmarking selective auditory attention in naturalistic listening.

Spatial semantic segmentation of sound scenes (S5) involves the detection and extraction of target sound events in audio files. However, in typical detection-extraction pipelines, errors in the detection module propagate to the extraction module, degrading overall performance. In this work, we develop a multichannel detection-extraction model and evaluate inference-time algorithms to improve the class detection and target sound extraction in a multi-stage setup. We evaluate two fitness functions: binary cross entropy and mixture consistency; two relabeling strategies: naive and conditional confusion matrix relabeling; and three source estimate evaluation models: a multi-label classifier, a fine-tuned single-source classifier, and an off-the-shelf audio judge. We thoroughly evaluate the impact of these design choices on the DCASE 2025 Task 4 dataset, and validate entropy-based gating of inference-time updates on the held-out evaluation set, laying the groundwork for inference-time updates of S5 systems in real-world scenarios.

We investigate noise disparity in multi-class speech classification tasks and develop a strategy to prevent classifiers from exploiting class-dependent noise characteristics. Building on our previous work for binary classification, we introduce a multi-class cross-augmentation scheme that exposes each class to the noise characteristics of the other classes, thereby removing the association between individual noise conditions and class labels. We compare this training-based approach with speech enhancement (SE) as a preprocessing strategy, which aims to suppress noise-related cues directly from the input. Experiments on multi-class emotion recognition show that cross-augmentation effectively mitigates the effect of noise disparity across a range of signal-to-noise-ratios, while SE has a detrimental effect on model performance.

Automatic dysarthric speech detection approaches can support traditional clinical diagnosis, which relies on costly and time-consuming evaluation by a speech and language pathologist. Existing automatic approaches predominantly rely on deep learning (DL). More recently, Large Audio Language Models (LALMs) have emerged as a promising alternative given their strong performance across various tasks, but their application to dysarthric speech detection has not yet been established. We propose a framework that fine-tunes LALMs for dysarthric speech detection on speech recordings combined with textual information comprising low-level acoustic features and speaker demographics. Across two LALMs, our framework outperforms DL-based baselines, with Qwen2-Audio-Instruct achieving state-of-the-art performance. An ablation study shows that incorporating acoustic features and speaker demographics during fine-tuning improves LALM performance, while LALMs alone exhibit only chance-level zero-shot performance. These findings establish an effective approach for adapting LALMs to dysarthric speech detection.

Knowledge-driven neural vocoders struggle to learn reliable fundamental frequency end-to-end, because spectral objectives provide weak supervision of periodic structure and lack phase information. We address this with a source-filter model whose alias-free additive source makes the instantaneous phase of the glottal cycle explicit; differentiating it yields the instantaneous frequency, and thus $F_0$, without an external tracker. Waveform error supervises only the deterministic harmonic path, while a spectral loss covers the full signal. On M4Singer and LM-SSD, the reconstruction is phase-aligned, reaching a signal-to-reconstruction-error ratio of 8.1 dB, while neural baselines remain negative. However, GOLF, given an external $F_0$, still reaches lower spectral distortion. On LM-SSD, the recovered $F_0$ attains the highest overall accuracy of any method tested, including supervised neural pitch trackers applied off the shelf, and the glottal closure instants come within 0.53 points of REAPER's identification rate, without any $F_0$ label.

Real-time conversational agents require speaker diarization that streams and runs on a CPU. Most systems apply an utterance-level speaker encoder to short, heavily overlapping chunks, which wastes computation and leaves the model optimized for the wrong task. We instead turn a state-of-the-art speaker recognition architecture into a causal frame-level encoder that reads the stream once and emits one embedding every 80~ms from a bounded two-second window of past audio, and pair it with online clustering that gates every update on the self-similarity of the stream. Trained only by distillation from an utterance-level teacher on simulated and out-of-domain mixtures, and evaluated with one fixed set of hyperparameters, the system is the most accurate streaming diarizer on low-overlap benchmarks at sub-second latency, degrades far less than cache-based systems as the number of speakers grows, and runs five times faster than real time on a single CPU thread.