audio software engineers

Abstract: Local music applications estimate key during library import and benefit from an analyzer that is small, fast, and runs in-process. We start from a compact chroma-template key analyzer that fits this budget but lags behind learned references. Learned systems reach higher accuracy, but their installed model and runtime dependencies can occupy tens to hundreds of megabytes. We present key lime, a 21.5k-parameter DSP-guided transformer reranker with a 115 KB model artifact and a 76.7 MB installed footprint, including the extracted ONNX Runtime. It adds 0.32 ms per track to deterministic DSP scoring. Trained from labels on 4,040 FMAK tracks, with 449 tracks for validation and 1,000 disjoint FMAKv2 tracks held out for evaluation, it improves MIREX from 61.08 to 73.55 +/- 0.46 and exact accuracy from 52.30 to 67.07 +/- 0.65 over three seeds (mean +/- population SD). It approaches a same-manifest local madmom KeyCNN reference at 74.16 MIREX. A 12-parameter circular input stage matches the accuracy range of a 468-parameter linear input projection.

Abstract: Large audio-language models (LALMs) frequently exhibit object hallucination, affirming absent sounds. The dominant mitigation, Audio-Aware Decoding (AAD), contrasts logits under real audio against an all-zero "blank" waveform, flagged by prior review as out-of-distribution. We introduce a black-box causal test that isolates the claimed evidence itself: localize a claim's evidence region via CLAP similarity, remove it by spectral gating, and compare against a random-window control across a 2x2x2 grid on three LALMs. On Qwen2-Audio-7B, localized ablation's raw AUROC trails blank's (0.613 vs. 0.721); this is a duration confound, not a localization failure, since blank removes the entire clip versus the grid's 2-4 s. Holding duration fixed, localization beats random removal at every cell on both models. Raw detection value remains model-dependent, especially on Audio Flamingo 3. The same evidence corrects decoding: contrastive decoding combining blank and language-prior contrasts, gated and ensembled, raises Qwen2-Audio accuracy from 57.1% to 71.9%, beating AAD alone. A unified causal-grounding profile does not transfer between models, indicating different grounding regimes. The source code and a detailed report of our findings, including identified limitations, are available at https://github.com/khalooei/causa.

Abstract: LLM-based text-to-speech (TTS) models can generate emotionally expressive speech, but how reference emotion is routed through the model and realized in decoded speech remains unclear. We introduce two emotion-sensitive metrics for matched neutral and emotional syntheses---a codec trajectory score and a late residual direction score---and use them to score activation-patching interventions. Under controlled matched-reference conditions, this analysis identifies a sparse source-to-readout component-level circuit: 23--27 attention heads and MLPs per emotion, roughly 5% of the components considered, recover or suppress 74--88% of the late emotion-readout shift on held-out cases. The circuit combines a shared component backbone with emotion-specific components; cross-emotion activation swaps reduce the target readout in 47 of 48 cases. In decoded speech, the same intervention produces consistent changes in pitch, energy, and spectral brightness over 24 matched pairs per emotion. A readout-matched residual-direction baseline produces only 17--27% of the intervention's pitch effect, showing that internal readout movement alone does not explain the decoded acoustic changes. These results trace a compact causal route from reference-derived prefix information to emotion-relevant properties of generated speech.

Abstract: Large Audio Language Models (LALMs) are prone to hallucinating and over-relying on text priors when simultaneously presented with audio and text inputs. To mitigate these hallucinations, we propose utilizing the multimodal Direct Preference Optimization (mDPO) objective, which forces the model to ground its generation in the acoustic input by contrasting intact and distorted audio counterparts. We extend this preference learning framework to the audio domain by applying a variety of acoustic perturbations. Evaluating the Qwen2-Audio backbone across the DCASE 2025 Challenge and AH Existence datasets, we demonstrate that extending mDPO to LALMs significantly enhances temporal reasoning in the complex DCASE dataset, and improves performance on basic existence verification in the AH benchmark. We identify temporal reversal, frequency masking, and random noise as the most effective perturbations. Ultimately, our approach achieves an absolute accuracy improvement of 14.0% on the DCASE 2025 dataset and 27.4% on AH Existence.

Abstract: A single compact encoder for both speech and music removes the need to maintain a separate model per domain. A practical recipe distils a speech teacher and a music teacher into two small students, then merges them into one model. Two merging algorithms exist: interpolating the student weights from a shared initialisation, or permuting their channels into alignment before averaging. Prior work fixed the distillation loss in both, leaving open whether the loss itself affects how well the students merge. We show that it does. Holding architecture and evaluation fixed, we replace the DistilHuBERT loss \Lkd{} with a correlation-based loss \Lcl{}. Under weight interpolation, \Lcl{} students stay closer to their own better endpoint in $17$ of $18$ comparisons and lead on the speech tasks at every interpolated weight. Under activation permutation, they need fewer channels rearranged at three layers, their matched channels correlate more strongly, and they again lead on speech. The two merging algorithms share no mechanism, yet both improve under the same loss substitution. This suggests that the gain lies in the representation \Lcl{} produces rather than in either algorithm.

Abstract: Large audio-language models (LALMs) perform strongly on individual audio tasks, but whether these capabilities can be reliably composed remains underexplored. We conduct a controlled diagnostic study of capability composition in LALMs, requiring models to integrate audio-attribute recognition, cue-conditioned segment selection, and downstream ASR or question answering. We construct two-utterance inputs with distinct acoustic cues to evaluate composition across environmental sound, gender, and emotion cues, with ASR, Math QA, and Factual QA as downstream tasks. Across four open-source LALMs, compositional QA accuracy decreases in 39 of 40 model-task-cue settings, by an average of 26.7 percentage points. ASR exhibits a similarly consistent degradation, with WER increasing in 39 of 40 settings by an average of 28.5 percentage points, while the magnitude of degradation varies across models, cue types, and cue salience. We further probe these failures through output format, positional preference, and chain-of-thought (CoT) analyses. Our study reveals a systematic gap between possessing individual audio capabilities and reliably composing them.

Abstract: Existing training-based speech emotion editing methods often require substantial task-specific training and can be unstable. This motivates us to investigate whether the pretrained generative dynamics of large-scale text-to-speech (TTS) models can be directly manipulated for training-free emotion editing. To answer this question, we probe the editability of pretrained flow-matching and hybrid TTS models by constructing a controlled test set and systematically diagnosing editing effects along the generative trajectory. Our analysis reveals that pretrained TTS models are substantially editable in emotion, but such editability is architecture- and trajectory-dependent and can be disrupted by early flow-matching steps, while cross-speaker emotion transport carries additional acoustic attributes beyond emotion. To address these limitations, we propose SEmoEdit, the first training-free framework that formulates emotion editing as dynamic velocity transport between source and target emotions, enabling robust, flow-based speech emotion editing directly within pretrained TTS models. SEmoEdit unifies three core operations: emotion replacement, emotion erasure, and continuous emotion interpolation, requiring neither parameter updates nor task-specific optimization. To systematically evaluate these capabilities, we introduce SEmoEditBench, a dataset comprising 600 editing cases, and conduct extensive experiments across state-of-the-art (SOTA) models and backbones. Our results show that SEmoEdit is highly effective and broadly applicable, outperforming existing training-based and activation-steering methods. Ultimately, this work reveals that pretrained speech flows possess rich, latent emotion-editing capabilities, providing useful guidance for real applications. Code, benchmark, and Audio samples are available at https://github.com/imxtx/SEmoEdit.

Abstract: This paper proposes an architecture for equipping large language models (LLMs) with audio-understanding capabilities without fine-tuning their weights. The proposed symbiotic architecture employs an injector module that writes audio-conditioned vectors directly into the target LLM's short-term memory, i.e., the key-value (KV) cache, enabling the LLM to behave as an audio language model (ALM). The architectural advantages are twofold. First, it improves the scalability of ALMs: because the proposed method bypasses the LLM during audio injection, the injection cost is governed by the injector width rather than the backbone width, and can therefore scale more slowly than the cost of full-backbone prefilling. Second, since the training scheme does not update the LLM weights, the original capabilities of the LLM are preserved without the risk of degradation from fine-tuning. The effectiveness of the proposed method is evaluated on both audio-understanding tasks (automatic speech recognition, audio question answering, and acoustic scene classification) and text-only tasks. We confirm that, while activating fewer parameters during audio prefilling, our architecture outperforms the conventional method with a frozen LLM and approaches the performance of a fine-tuned ALM, all while preserving the backbone LLM's original text-only task performance by construction.

Abstract: Modern neural audio watermarking systems typically embed a message repeatedly across time and then collapse the resulting temporal evidence into a single payload using averaging, voting, or another fixed aggregation rule. We argue that this temporal collapse limits both robustness and the recovery of multiple payloads, and that the limitation can be addressed without retraining the underlying watermarker. We freeze a pretrained watermarker's encoder and detector and train only a low-latency Conformer-based decoder. The decoder consumes the detector's temporal soft outputs, which a system-specific adapter pools into a sequence of window-level representations, and predicts the embedded message. On three frozen watermarkers (AURA, AudioSeal, and WavMark), the learned decoder improves recovery of attacked messages and yields higher detection AUROC point estimates on all three. Under controlled full- and partial-coverage multiplexing, it improves joint-exact recovery of two alternating payload words by 9.7-48.0, 6.9-17.3, and 8.2-14.2 percentage points, respectively, under one to three chained attacks on feasible clips.

Abstract: A finite-difference bowed-string model with implicitly resolved Stribeck friction is presented, with a regime diagnostic, the Schelleng bow-force limits on four strings, and a comparison of learned bow controllers. Implicit resolution is necessary, and quantitatively so: a lagged contact force cannot capture the string on a discrete grid, so no stick phase forms at any bow force. With friction, impedance and quality factor taken from published measurement rather than fitted, all four strings return a stick fraction of 89.1% against an ideal 90%. Schelleng's maximum bow force is recovered on every string. The minimum is not: it follows $Z v_b β^{-1}$ rather than the predicted $Z^2 v_b β^{-2}$, reducing both squared dependences to first powers. Six controllers at matched capacity, over four strings and twenty seeds each, place a gated recurrent network ahead of a feedforward one, by most under a mid-stroke disturbance. The feedforward network completes more strokes only from a start the model's own playability map places outside the Helmholtz region. A minimal gated variant fails because gates computed from the input alone cannot clear a latched state. Training loss selects neither the capacity nor the context length, and no learned controller improves on the lookup rule that generated its labels. That bound has a domain. Regressing the controller's score on the rule's gives a slope of 0.32, more than ten standard errors below unity, so the controller overtakes the rule where the rule fails and is bounded by it where it holds. Under a rigid finger stop the plant is provably invariant, so transfer loss between pitches belongs to the controller alone and is traced to one feature. A regime classifier without a stick test labels small-amplitude periodic slipping as Helmholtz motion, and a harmonicity measure rates a string the bow never grips above Helmholtz motion.