Papers

Non-verbal vocalizations (NVVs) carry important paralinguistic information but are often omitted by conventional automatic speech recognition (ASR) systems. The ISCSLP NVVSpeech Challenge requires joint transcription of lexical content and 16 NVV categories under limited and highly imbalanced supervision. We present a data-centric NVV-aware ASR pipeline based on cross-dataset label harmonization and a two-stage sampling schedule. We map heterogeneous source labels to the official taxonomy and exclude samples without a reliable mapping. Our schedule first uses square-root category sampling to moderate the long-tailed distribution and then applies uniform-category fine-tuning. On a fixed local validation split, square-root category sampling performs best among the tested single-stage settings. The final two-stage system obtains an official score of 63.86 and ranks fourth in Track 1.

Voice anonymization aims to protect privacy by suppressing speaker identity while preserving linguistic content and prosody. However, residual speaker attributes in non-identity representations may still increase linkability and weaken privacy protection. To this end, this paper proposes a fine-tuning strategy with a pinhole loss for well-trained voice anonymization frameworks to further reduce residual speaker attributes. Inspired by the pinhole effect, the pinhole loss measures the linkability of anonymized utterances from the same source speaker. By minimizing this loss, linkability is reduced, thereby improving privacy protection. Experiments on multiple anonymization frameworks, pseudo-speaker generation methods, and datasets show improved privacy protection while maintaining utility. Audio samples can be found in https://anonymous.4open.science/r/Pinhole-loss-fine-tunning-4628.

Conventional general sound recognition systems typically output deterministic sound event labels, implicitly assuming that the target sound class can be correctly identified from the input audio. However, in real listening situations, the sound event class is not always clearly identifiable. Human listeners may nevertheless understand their surroundings from an ambiguous sound without identifying its exact sound event class. This motivates a discussion of how the outputs of sound recognition systems should be redesigned under such uncertainty. As a basis for this discussion, this paper proposes an output representation for sound event recognition that combines a sound event class, its confidence score, and an onomatopoeic description of the sound. The proposed representation preserves conventional class-based recognition while providing an additional onomatopoeic description of acoustic characteristics that can remain informative even when the class prediction is uncertain. Experiments using ESC-50 and ESC-50-Onomatopoeia show that the proposed method achieves sound recognition performance comparable to that of a conventional recognition-only system. In addition, an LLM-as-a-judge evaluation and subjective listening experiments indicate that the proposed output is preferred over conventional deterministic outputs based on the sound event label, particularly when used to support understanding of the surrounding environment. These results suggest that such output representations can make sound event recognition more informative and communicative under uncertainty.

Pitch estimation on an edge device is constrained in three ways at once. The model must be small, it must stay accurate when the input is noisy, and one frame must be produced inside the frame period. In this report the Frequency Convolution Network (FrCN) of our earlier work is factored into a low-rank form. The number of parameters is reduced by 35.9%, from 17,787 to 11,397, and accuracy is not reduced, either in domain or on two corpora the model was never trained on. The range of the noise used during training is also shown to dominate the architecture in setting how the model behaves when the interference is severe. When the training noise floor is lowered from +6.02 dB to -20 dB, 0.35 points of clean RPA50 are lost and RPA50 at -20 dB is raised from 2.44 to 22.41. This effect is about two orders of magnitude larger than any architectural effect that was measured. The semi-orthogonal constraint used in TDNN-F is found to be redundant with the normalization inside the bottleneck, and accuracy is reduced when both are applied. Whether the factorization saves time depends on the runtime: in eager PyTorch the factored model is 45% slower, while in a compiled kernel it is 15% faster. For deployment, a small C kernel was written. It needs 83 kB on disk and no runtime library beyond libc and libm. It is faster than OpenBLAS on all four CPUs that were tested, faster than ONNX Runtime by 3.8 times, and faster than PyTorch by 13 times. Its output was checked against PyTorch on 271,893 held-out frames per model, and the same pitch bin was selected on every one of them.

Hands-free communication devices and speakerphones are inherently affected by acoustic echo and background noise. To mitigate these impairments, end-to-end discriminatively trained neural networks have emerged as the best-performing approach in research and deployment. While recent advancements in generative methods have provided remarkable results for various speech enhancement tasks, diffusion-based acoustic echo control (AEC) research is still restricted to non-causal, utterance-level processing, thereby not widely applicable in practice. With this work, we are the first to propose low-delay (i.e., causal) diffusion-based joint AEC and noise control models DiffVQE2 / DiffVQE2-S, excelling the so-far state of the art DeepVQE / DeepVQE-S models in multiple objective metrics, and, most importantly, in subjective MOS, respectively. In addition, our models are less complex. Furthermore, we show that a limited lookahead applied to the efficient DiffVQE2-S model allows for an even higher performance. These results have been obtained on the ICASSP 2023 AEC Challenge blind test set. We claim the first streaming-capable, diffusion-based acoustic echo and noise control that excels state-of-the-art discriminative approaches.

A microphone on the airframe of a small multi-rotor UAV is dominated by rotor ego-noise, so spoken flight commands arrive at negative signal-to-noise ratio (SNR). We study small-footprint keyword spotting (KWS) for a ten-word command vocabulary under real ego-noise, training on one quadrotor and testing on another. Besides per-clip accuracy we measure the streaming false-alarm rate on 4.4 h of continuous rotor noise. We compare classical noise-robust front-ends (CMN, PCEN, spectral subtraction), test-time adaptation, and two front-ends that track the per-band ego-noise floor over the two seconds preceding the decision window and either subtract it (normalisation) or feed it to the network as a second input channel (conditioning). Per clip, all front-ends look alike: +2-3 points on average, up to +14 at -15 dB. On continuous rotor noise they differ sharply. Normalising front-ends fire about ten times more often than plain log-mel at the same threshold and end up below it at a budget of one false alarm per hour (-9 points at 0 dB). Conditioning keeps the baseline's false-alarm rate and turns its gain into detections (+8 points at -10 dB over three seeds); on top of PCEN it gives the best per-clip accuracy and false-alarm rate, and a level-anchored variant is also invariant to the microphone gain. Real drone+interferer recordings expose the remaining failure mode, environmental sounds and bystander speech, which training negatives halve. A single script reproduces all on-device numbers on the NVIDIA Jetson Orin NX flight computer, where the complete pipeline costs 3 ms per 100 ms hop on the CPU.

Vinyl restoration systems must distinguish isolated clicks, short bursts, dense crackle, and overlapping damage before selecting a repair operation. We present a lightweight two-stage detector in which signal-informed sparse, burst, and dense experts produce complementary defect evidence, and a temporal backend converts that evidence into stable repair regimes. The backend factors the five-way decision hierarchically, applies a validation-only mixed regime gate, and decodes with validation-selected transition penalties that outperform a maximum-likelihood transition matrix on the same emissions. On a source-separated synthetic benchmark of 597 non-overlapping 15 s excerpts drawn from 21 recordings, the held-out system reaches 0.730 pooled five-class macro F1 (95% CI 0.694-0.761), an improvement of 0.043 over a flat frame-level router (paired bootstrap p=0.001). Clean and dense frames are highly reliable, while sparse, burst, and mixed frames remain far more ambiguous.

Recent advances in Speech Language Models (SpeechLMs), which integrate large language models with speech foundation models, have enabled unified sequence modeling of speech processing tasks. However, many SpeechLM-based approaches to speaker diarization (SD) are tightly coupled with automatic speech recognition (ASR) and evaluated using word-level metrics, making it difficult to assess SD performance independent of ASR accuracy. In this work, we investigate ESPnet-SpeechLM as a token-based backbone for generating SD hypotheses, formulating SD as autoregressive generation of structured tokens conditioned on acoustic input. We systematically compare two output representations: an event-based representation that explicitly models speaker turn onset and offset timestamps, and a frame-based representation that predicts frame-level speaker activity. To provide structured conversational cues, we further incorporate auxiliary tasks including speech activity detection, overlapped speech detection, and speaker turn counting within the output sequence. Across multiple meeting datasets, we find that event-based representations produce more stable and consistent SD outputs than frame-based representations. Our analysis shows that outputs generated by SpeechLMs encode useful temporal SD structure, but full-meeting SD remains limited by recording-level speaker tracking and overlap-related misses. Explicit speaker-linking post-processing substantially reduces speaker confusion, suggesting that robust SpeechLM-based SD requires persistent speaker tracking and overlap-aware generation.

Most neural speech enhancement systems use a fixed processing depth for all inputs, which can introduce unnecessary computation when fewer refinement steps are sufficient. We propose Adaptive Depth and Expert Refinement (ADER), a parameter-shared progressive enhancement framework with input-dependent computation. ADER combines an Adaptive Depth Controller (ADC) for hard early termination with a Conditional Expert Router (CER) that selects one lightweight residual adapter at each executed refinement iteration. We further introduce Exit-aware Intermediate Supervision (EIS) to directly optimize candidate intermediate outputs for early exit. On VCTK-DEMAND, ADER reduces the parameter count and average computation of MP-SENet by 70.4% and 51.3%, respectively, while achieving a WB-PESQ of 3.37. Overall, ADER enables input-dependent refinement and reduces redundant computation during inference.

Recent advances in Audio LLMs have achieved human-level speech recognition, yet existing systems struggle to capture paralinguistic aspects such as speaker traits, expressive variations, and environmental acoustic conditions. To address this, we design a framework of 22 paralinguistic characteristics and create a dataset of over 1.2M Audio-QA pairs. We develop ParA-LLM, trained with a two-stage curriculum: first on single-attribute questions to build foundational knowledge, then on multi-attribute questions for joint reasoning over speaker and acoustic characteristics. We also release ParA-Bench, a benchmark of 6,000 multiple-choice questions across speaker-speech, acoustic, and mixed categories, where frontier models like GPT-4o-Audio achieve only 36% accuracy. ParA-LLM surpasses state-of-the-art Audio LLMs like GPT-4o-Audio by 7.5% on ParA-Bench, with additional gains of 1.13% on MMAU-Pro Speech and 7.49% on MMAR Speech.

Recently, text-to-speech systems have made significant progress in speech expressiveness and controllability. However, the speaking style of generated speech typically relies on clear user-specified instructions. In natural conversations, speaking style should be naturally inferred from the preceding conversational context. Therefore, we propose COT-TTS, a context-aware, reasoning-based text-to-speech task. Given historical conversation audio, target text, and a reference speech, the system should comprehend the conversational context, infer an explicit intermediate reasoning, and finally synthesize the target speech with the specified timbre. To support this task, we constructed a large-scale bilingual conversational speech dataset comprising 9 million training samples, including a high-quality subset of 1 million samples. We further constructed a source-disjoint benchmark with 800 human-verified samples and established strong task-specific baselines. Additionally, we developed end-to-end autoregressive models with parameter sizes of 0.6B and 1.7B, generating emotion-labeled transcripts, editable speech style inferences, and speech tokens. Experimental results show that the proposed model achieves performance comparable to large-scale baseline systems with significantly fewer parameters. At the same time, the model performs well in terms of duration consistency and emotional consistency, and can generate appropriate emotional, stress, and rhythmic variations based on the conversational context. To facilitate future research, we will publicly release the data construction pipeline, dataset, trained models, and related resources. The demo page and additional resources are available at https://luckybian.github.io/COT-TTS

Accent conversion is the task of modifying a speech recording so that it sounds closer to a target accent while preserving linguistic content and other speaker-related characteristics. Most accent conversion systems use trained, generative models. Although they can induce target-accented speech, the strength of accent modification is not controllable at inference time, making it difficult to analyse how the strength of accent conversion affects source preservation. We propose Partial Accent-Control Editing (PACE), an accent conversion framework based upon the editing of frozen WavLM representations without the training of an accent-conditioned generator. A constrained edit is first applied to source WavLM features, which are then fused with target-accent reference features retrieved from non-parallel accent examples. Fusion weights are used to control the trade-off between accent conversion strength and the degradation of other source attributes. Using a suite of five metrics, and with the cost of accent-entangled speaker similarity, we show that PACE is substantially superior to a pair of competitive baselines in terms of both accent conversion and source preservation.

Spatial room impulse responses (SRIRs) capture directional distributions of acoustic sound-sources and their reflections. However, collecting SRIRs of moving sound-sources remains a challenge, requiring complex interpolations across measurements that account for multi-path spatial-temporal dynamics. This paper investigates the Wasserstein metric and displacement for evaluating interpolated SRIR echo densities in the spherical harmonic domain. We present novel sum-of-magnitude square expansions for efficiently fitting probability density functions, maximizing likelihood, inverse sampling, and computing spherical sliced Wasserstein interpolations. Experiments compare the Wasserstein displacements and metric to linear and geometric interpolations of SRIR image-source densities on a line-path, and demonstrate model-order reduction.

Speech enhancement (SE) models degrade under domain shifts and have to adapt to unseen target domains during deployment. Most existing test-time adaptation (TTA) methods for SE do so by adapting a subset of the model weights using a self-supervised loss, which requires backpropagation at test-time and permanently alters the model. We instead recalibrate the prediction itself and propose BLINC, a training-free TTA method that remaps the predicted time-frequency mask onto a bimodal target distribution by histogram matching. At test-time, the target distribution is parameterized from blind features of the noisy recording, so neither a reference distribution from a classical algorithm nor online metric optimization is involved. BLINC improves the overall quality of both evaluated SE models on almost every target condition and matches or exceeds the loss-based TTA baselines at minimal overhead.

Long transcripts are costly inputs for downstream NLP systems and often contain irrelevant context. We study query-conditioned topic localization: predicting the sentence span in a transcript that best addresses a topic-title query. To improve span localization, we reuse ASR encoder states as sentence-level representations and fuse them with textual embeddings. This lets lightweight span locators exploit speech information without running a separate audio encoder. Experiments on two public datasets show consistent gains over text-only baselines, especially under strict boundary-matching criteria. Cross-dataset experiments further indicate that the benefits are strongest for structured or semi-structured speech, while gains on spontaneous speech are limited and mixed.

Modeling stiff string vibrations is challenging due to their dispersive and high-frequency characteristics. This study investigates the effectiveness of Physics-Informed Neural Networks (PINNs) in simulating the transverse vibration of a one-dimensional linear stiff string with sharp initial conditions induced by plucking. The governing Partial Differential Equation (PDE), along with the associated initial and boundary conditions, is incorporated directly into the loss function of the neural network. For the reference measurements, a wire-breaking experiment was performed to excite the string, and its vibration response was captured using a laser profiler. The comparison between experimental measurements, finite-difference time-domain (FDTD) simulations, and PINN-based simulations shows good overall agreement, highlighting the potential of PINNs for modeling stiff-string vibrations.

We present the Spoken Wikipedia Presentation Corpus, an extension of the Spoken Wikipedia Corpora featuring LLM-generated slide decks for multimodal ASR. Slides are created from LLM-segmented sections using a hybrid pipeline that combines LLM-based content planning with rule-based design decisions. For each section, an LLM generates a slide title, bullet points, a takeaway message, and a visual description that is used to create an illustration. Rule-based matching then selects layouts, themes, and styles to produce the final slides. A vision LLM extracts slide text as Markdown. We evaluate multiple ASR and spoken language models (SLMs). The best model achieves an average micro-WER of 10.23% and an average micro-CER of 6.48% on audio-only inputs. English yields the lowest error rates, followed by German and Dutch, while performance declines across lower-resource languages. Although audio-only baselines are strong, multimodal zero-shot prompting of omni models remains challenging. The aligned slide, text, and audio data show a strong potential to improve recognition through cross-modal context.

The success of Large Audio Language Models has driven the development of massive multimodal networks exceeding billions of parameters. However, the demand for privacy-preserving, low-latency processing has shifted focus toward Small Audio Language Models (SALMs) capable of on-device execution. In this paper, we introduce Samsone, a family of SALMs designed for edge computing. Our core model, Samsone-134M, establishes a new state-of-the-art for its size class across multiple benchmarks. We further explore the scaling laws of SALMs by introducing Samsone-99M and Samsone-356M. Despite their compact footprint, the Samsone family delivers performance competitive with models orders of magnitude larger. To foster open research and reproducibility, we train Samsone on publicly available data. We release the training code, model weights, mobile-optimized checkpoints and provide an open-source Android application to demonstrate real-time on-device inference of Samsone.

Tabla Stroke Transcription (TST) converts tabla audio into symbolic stroke sequences, but the scarcity of annotated recordings makes fully supervised training challenging. We propose a semi-supervised framework that uses sequence-level labelled and unlabelled tabla recordings. A teacher model generates pseudo-label sequences, while a Stroke-Level Confidence Estimation Model (S-CEM) estimates confidence for each predicted stroke. To improve training with uncertain pseudo-labels, we propose Confidence-Guided Markov Weighted Alternative Temporal Classification (CMW-ATC), which weights candidate sequences within uncertain spans using learned stroke transitions and reliable neighbouring strokes. Experiments across three TST evaluation settings show consistent improvements over conventional teacher–student training and Alternative Temporal Classification in its replacement form (ATC-R). Ablation studies further show the contributions of S-CEM, Markov transition weighting, and the use of reliable strokes on both sides of uncertain spans.

Audio deepfake forensics is moving beyond a simple real-or-fake verdict toward attribution: which system generated the audio clip? Most source-attribution methods cast this as closed-set classification, so they cannot name a generator that was absent from training, a gap that widens with every newly released text-to-speech (TTS) system. We instead frame attribution as cross-modal retrieval: each generator is described in natural language, and a clip is attributed by retrieving the description closest to it in a shared audio-text embedding space. Adding a new system then takes nothing more than writing its description, with no retraining and no new classifier head. Our model couples a frozen Wav2Vec2-BERT audio encoder with a frozen E5 text encoder and aligns them through small trainable projection heads, using a contrastive objective that combines a cross-modal supervised-contrastive loss with intra-modal terms. We evaluate on MLAAD v9 (140 TTS models, 51 languages) under 10-fold leave-models-out cross-validation. For generators it has never encountered before, the model reaches a model-level mean reciprocal rank (MRR) of 58.4%. Even when the correct model is not identified, the audio clip is often matched to systems that share the true generator's vocoder, acoustic model, or architecture. Because the same embedding space also answers natural-language attribute queries, one set of descriptions covers both open-set attribution and attribute-level forensic profiling.

We define OmniVChat (Omni Video Chat) as the task of native audio-visual dialogue between a user and an omni model. In OmniVChat, omni models directly and simultaneously receive audio and video from a user and return text. The user's query is embedded in the audio and video, without a separate text question, external captioning, or speech recognition. Direct audio-visual input reduces external latency and computation while preserving perceptual cues. However, research on OmniVChat faces two constraints: data availability and evaluation. Recordings of people using their own devices are scarce. Furthermore, good replies often depend on multimodal context and can be phrased in many ways, making keyword matching unreliable for evaluation. Recent progress in agent systems and video generation makes generation for comprehension viable, which means using synthesized dialogues for training and evaluation. Therefore, we present OmniVChat-Studio, a multi-agent data engine for synthesizing single- and multi-turn audio-visual dialogues. We use synthesized dialogues to build OmniVChat-Bench, an evaluation benchmark that evaluates omni models' basic dialogue abilities across five ability categories. Replies are judged by a large language model based on explicit scoring criteria. We also present OmniVChat-RL, a reinforcement learning reward design that jointly targets reply correctness, efficiency, and style in OmniVChat. Training Qwen3-Omni-Instruct with OmniVChat-RL on synthesized dialogues improves its performance on both OmniVChat-Bench and the human-recorded OmniVChat-Bench-Human. These gains validate the reward design and show transfer to real-world dialogues in training and evaluation.

Automatic speech recognition models suffer from catastrophic forgetting when adapted to new domains, accents, or downstream tasks. This problem becomes increasingly important with the growing use of speech foundation models, where adaptation should be both memory-efficient and safe, preserving the broad capabilities learned during pretraining. Gradient Low-Rank Projection (GaLore) has recently been proposed as a memory-efficient fine-tuning method that keeps model parameters full-rank while reducing the optimizer memory through low-rank gradient projection. However, GaLore does not account for catastrophic forgetting. We propose Curvature-Guided GaLore (CGaLore), which incorporates old-task curvature information when selecting the low-rank projection bases. Specifically, CGaLore filters current-task gradients using Kronecker-factored approximate curvature from previous tasks before computing the gradient subspace. Our experiments show that CGaLore enables effective adaptation while alleviating forgetting, outperforming state-of-the-art continual learning baselines. Extensive ablation studies confirm the practical applicability of CGaLore.

Discrete audio tokenization has become the critical interface between raw waveforms and autoregressive modeling in recent music generation. As a result, music tokenizers must simultaneously support high-fidelity reconstruction and produce discrete sequences that remain amenable to language modeling. Existing reconstruction-oriented tokenizers often mix musical structure with fine acoustic details, producing high-entropy tokens that are hard to model. In contrast, semantics-guided alternatives are designed for speech and do not fit music well, often hurting reconstruction quality. We address these trade-offs by rethinking music tokenization around a measurable notion of music semantic content grounded in downstream Music Information Retrieval tasks. Guided by this definition, we propose MuSeC, a music semantic codec that factorizes semantic and acoustic content directly from mixed signals without source separation. MuSeC preserves information required for high-fidelity reconstruction while producing more LM-friendly discrete units. Empirically, it improves reconstruction quality and yields more predictable token sequences, providing a practical foundation toward high-fidelity LLM music generation. Demos are available at https://longwaytog0.github.io/MuSeC/.

Recent speech enhancement systems combine self-attention and Mamba to capture global interactions and long-range dependencies. Yet these hybrids usually operate as sequence mixers and do not explicitly exploit harmonic periodicity, a strong cue for preserving voiced speech under noise. Perceptual optimization poses another challenge. PESQ is non-differentiable, so many methods train auxiliary metric discriminators that increase complexity and introduce adversarial instability. We propose \ours, a harmonic-aware and discriminator-free speech enhancer built around two components. (i) The Time-Frequency Harmonic-aware Attention-Mamba (TF-HAM) block runs self-attention and bidirectional Mamba in parallel along both spectrogram axes, then applies a speech-adapted autocorrelation feed-forward network to encode local periodic structure. (ii) Metric-explicit perceptual refinement (MEPR) combines differentiable PESQ and log-likelihood-ratio losses to expose perceptual metric structure without a learned surrogate. On VoiceBank+DEMAND, \ours achieves 3.69 PESQ and 4.41 COVL with only 2.39\,M parameters, outperforming or matching discriminator-based systems. Inference-time perceptual contrast stretching further raises PESQ to 3.79 without retraining. The source code will be available at https://github.com/shangfuu/HAMMER.git.

Modern automatic speech recognition (ASR) systems trained on extremely large datasets can produce transcripts with numbers written in Arabic numerals. This creates a need for fair comparison with models that output verbatim texts and proper processing of reference transcripts. Popular approaches often reduce text normalization to lowercase and remove punctuation, with no additional normalization applied to languages other than English. In this work, we analyze the impact of normalization of numerical expressions in the evaluation of ASR systems in various languages, using Polish as an example of a highly inflective language. We perform experiments on VoxPopuli and The Polish Parliamentary speech datasets and estimate word error rate (WER) differences for different text normalization approaches. We show that the difference due to the lack of number normalization in WER may be substantial - more than 2 percentage points, and often higher than the differences between systems in popular multilingual benchmarks.

Recent work revealed a striking Clever Hans effect in the Pitt dataset, where Alzheimer's detection achieved nearly 100% accuracy using only silent audio segments. This raises serious concerns about hidden confounding factors in speech-based health datasets. We systematically investigate whether similar biases exist across five widely used clinical speech corpora: DAIC-WoZ (depression), TORGO (dysarthria), Neurovoz (PD), MDVR-KCL (PD), and UCLASS (stuttering). For each dataset, we compare classification using the first second of audio, silent segments, and full recordings, and evaluate both raw and denoised signals. Across all datasets, silence-only classification frequently matched or exceeded full-audio performance, suggesting that classification performance may be influenced by dataset-specific confounds in addition to disorder-related speech characteristics. These findings question the reliability and generalisability of speech-based biomarkers and call for stricter methodological reporting, preprocessing transparency, and bias mitigation.

Natural spoken interaction requires more than streaming ASR, language generation, and speech synthesis: a system must react to overlap without canceling on every acknowledgment, prepare a response before a turn is certain, and ensure canceled audio cannot enter conversation history. We present Voice-Light, a full-duplex cascaded voice agent that combines immediate acoustic onset, a causal adapter sharing a streaming ASR encoder, reversible playback control, and private speculative response generation. Structured tool calls execute concurrently with audible bridge speech, while browser acknowledgments make rendered audio authoritative for durable history. Locked evaluation on 1,673 real-conversation silence candidates found that an earlier learned completion checkpoint preserved a 2.70% false-cutoff rate but reached only 12.53% end-of-turn recall, compared with 95.60% for a Silero timing policy. The deployed system therefore retains a hybrid controller rather than claiming a learned-policy replacement. Across three unscripted operator-run microphone sessions, 36 measured response turns had a 758 ms median from final VAD endpoint to first server audio; 21 turns were below 800 ms. These sessions are an instrumented case study, not a controlled user evaluation. We release the synthetic data, model artifacts, evaluation code and summaries, source code, and deployment configuration supporting the result.

Computational models of auditory physiology commonly target specific responses or stages of the auditory pathway, limiting their ability to integrate findings across experimental paradigms and neural timescales. We present a foundation model of human auditory electrophysiology: a causal neural network trained to map binaural acoustic waveforms directly to high-sample-rate EEG. The model was trained on approximately 250 hours of EEG data from 92 subjects, with varied electrode montages and stimuli spanning tonebursts, speech, and music. We tested whether the model recovered effects of stimulus rate, frequency, and presentation method on auditory brainstem responses (ABRs); subcortical and cortical temporal response functions (TRFs) to continuous speech; and the click-evoked binaural interaction component (BIC). Predicted ABRs and TRFs reproduced established response morphology and stimulus-dependent effects, with model-grand-average correlations falling within the corresponding subject-level human distributions. The model-predicted BIC metrics closely resembled the values reported in the literature. These findings demonstrate that a single audio-to-EEG model can capture auditory physiology across paradigms and timescales, supporting future in silico experimentation and hearing technology applications.

FarmerChat is Digital Green's AI-powered agricultural advisory assistant for smallholder farmers, who access it in their own language through text, voice, or photographs. Voice is a critical channel for this population, yet field-recorded speech is challenging for general-purpose automatic speech recognition (ASR) because recordings frequently contain machinery noise, background media, competing speakers, and domain-specific agricultural vocabulary. These conditions disproportionately affect crop, pest, chemical, and quantity terms that carry the meaning of a farmer's query. We present a modular, model-agnostic pipeline for improving ASR quality in FarmerChat without fine-tuning or replacing the underlying ASR model. The pipeline combines gated audio enhancement, speaker diarization and target-speaker selection, ASR, domain-aware correction using a weighted agricultural lexicon, and a quality gate for detecting unreliable transcripts. Only the diarization stage is fine-tuned; all other stages use off-the-shelf models behind common interfaces. We evaluate the pipeline on human-annotated FarmerChat recordings in Hindi, Telugu, and Odia using word error rate (WER) and a domain-weighted error rate that gives greater importance to agricultural terminology. The largest improvements occur on multi-speaker recordings, where target-speaker selection prevents competing speech from entering the transcript. Across the full corpus, the pipeline reduces WER by 16-23% relative on three cloud ASR models and by 5% on an on-device model. On multi-speaker recordings, the reductions are 32-42% for the cloud models and 16% for the on-device model. All reported reductions are statistically significant. These results show that targeted preprocessing, speaker selection, and domain-aware post-processing can substantially improve agricultural speech transcription while preserving the underlying ASR model.

Streaming target speaker extraction must maintain a representation of whom to extract while the target may fall silent, be masked by interference, or drift acoustically away from enrollment. Existing systems typically hold this state as a stored embedding updated by hand-designed rules. Across 22 configurations, including confidence-gated and oracle-activity-gated updates, we show that this family lies on a stability-plasticity frontier: even perfect target-activity information cannot combine robustness to target absence with adaptation to enrollment-mixture mismatch. We therefore meta-train speaker-state dynamics through the closed streaming loop, exposing the updater to its own contaminated evidence. Our proposed 41k-parameter anchored fast-weights (AFW) memory moves beyond the measured heuristic frontier, gaining 3.0 dB over the best heuristic under severe mismatch while staying within 0.9 dB of static enrollment after 30 s of absence, at under 5% runtime overhead. A gated recurrent unit (GRU) control confirms that the gain is not AFW-specific, while AFW is smaller and more interpretable: under severe mismatch, its write residual grows and aligns with the target rather than the interferer. Code is publicly available at https://github.com/ym2976/anchor-fast-weight.