arXane › Electrical Engineering and Systems Science › Electrical Engineering and Systems Science
Wednesday, 4 November 2026: 4 new papers, 2 cross-lists and 6 replacements in Audio and Speech Processing (eess.AS), in arXiv's order, with full abstracts. The same papers as arXiv's own listing, set as a newspaper; every title links to arXiv.
Nepali TTS has reached a point where multiple systems are available, but there is no shared evaluation to compare them. NepTTS-Bench was created to address what existing automated and manual evaluation systems lack for Nepali TTS. A controlled benchmark with phonological grounding which contains 205 sentences (77 minimal pairs targeting aspiration, retroflex/dental contrasts, nasalization, gemination, and schwa deletion) which is rated by 180 speakers. There are 6,962 ratings for 10 TTS configurations and two human references. Along with this 7 ASR systems were evaluated for round-trip CER. SCOREQ was also run as an automated baseline. The key finding is that automated evaluation like SCOREQ prefers proprietary systems like ElevenLabs but native listeners prefer Piper. This is contributed by the fact that most of the systems not properly tuned for Nepali tend to show Hindi phonological behaviour and our minimal pairs are designed to surface such contrasts. We also present NepaliMOS predictor which is fine tuned on collected ratings. NepaliMOS achieves system-level rho = 0.90 on the dense panel but drops to utterance-level rho = 0.43 under leave-one-system-out evaluation. This is statistically significant and confirms that NepaliMOS captures genuine signal but cross-provider transfer is limited. A human calibration is recommended for new systems for which we release a Python evaluation package for practitioners to run the benchmark against new systems.
Speaker-aware applications rely on enrolled voice profiles, yet collecting enrollment speech from every participant is impractical. Meeting archives provide an alternative source of enrollment speech, comprising recordings of recurring participants paired with expected attendance lists. We investigate whether expected attendance lists can be cross-referenced with diarized anonymous speaker clusters to identify reliable enrollment speech for recurring participants. The resulting task, multi-participant automatic enrollment, must accommodate two types of attendance noise: a listed participant may not speak, and a speaker may be unlisted. To this end, we introduce a benchmark for the task: a synthetic corpus calibrated to an in-house meeting corpus, paired with a contamination-sensitive evaluation protocol. The reference method we introduce scores cross-session cluster chains against expected-attendance patterns and abstains when the evidence is insufficient. We quantify the reference method's benchmark performance as a trade-off between enrollment success and contamination. Applying synthetic-development thresholds unchanged to real meetings exposes a synthetic-to-real gap that remains an open challenge. We will release the benchmark generator and evaluation code.
A speech deepfake detector estimates whether a recording is synthetic, but a forensic analyst also asks which system produced it, whether recordings share a source, and what evidence supports each conclusion. We present VeriSpoof, an audio language model that unifies five tasks with shared weights: deepfake detection, generator attribution (generator, acoustic model and vocoder), recording-condition analysis, synthesized-span localisation, and source verification. Each decision is supported by grounded rationales whose statements can be independently verified. Answers are scored by sentence likelihood, allowing detection and source verification to be thresholded without separate heads. Rationale statements can be checked against waveform measurements, span labels, or generator metadata. Counterfactual examples train decisions to follow the stated evidence. Across the 14 Speech-DF-Arena test sets, VeriSpoof-Plus achieves 2.14% pooled EER, ranking third of 28 systems and first among open-weight systems. VeriSpoof-Base, trained only on official training splits of eight corpora, achieves 4.43% EER, less than half that of the next open-weight detector. With the same weights, VeriSpoof-Plus verifies whether recordings from generators unseen in training share a source at 6.2% EER. Controlled studies show that the audio representation governs detection, the LLM governs recording comparison, and source supervision is necessary for neural-codec fakes. Statement-checkable rationales catch partial fakes missed by one-step answers (99.6% vs. 78.4% on PartialEdit) and determine decisions under edits unseen in training (99.5%). Training updates only 29.7M of 2.85B parameters, fitting a single 24 GB GPU with gradient checkpointing. We release both models, inference code, a demo, and VeriSpoof-Bench, the evaluation suite covering these questions.
Generative speech models synthesize audio that is difficult to distinguish from human recordings. Synthetic speech can be watermarked as a proactive disclosure method, but watermarking models remain vulnerable to resynthesis attacks using neural audio codecs. This article investigates speech watermarking robustness across neural codec and digital signal processing attacks, when specific neural codecs are used for channel augmentation during training. We also present a novel zero-bit watermarking system that uses Descript Audio Codec as the generator. Our results demonstrate that the choice of augmentation codec significantly influences robustness generalization. Experiments show that augmentation with low-bitrate, high-quality neural codecs-Mimi and SNAC-results in promising robustness against unseen neural codecs and traditional attacks with limited quality impact.
Respiratory sound classification from auscultation recordings is confounded by heterogeneous acquisition conditions, including recording device, auscultation site, and patient characteristics, that introduce systematic variability unrelated to the underlying pathology. Although such metadata is often available, existing approaches either ignore it or rely on na\"ive fusion strategies that do not model how distinct metadata sources influence different components of the acoustic representation. We propose a family of metadata-conditioning mechanisms for the Audio Spectrogram Transformer (AST), progressing from gated residual fusion to Feature-wise Linear Modulation (FiLM) adapted to the Transformer architecture. We further introduce two novel variants: Token-Aware FiLM (TAFiLM), which applies distinct modulation to classification and patch tokens, and Soft-Factorized FiLM (SoftFiLM), which learns soft masks to align metadata factors with disentangled feature subspaces. Experiments on the ICBHI benchmark demonstrate consistent gains across all conditioning strategies. SoftFiLM achieves 64.11% on the official 4-class task (+3.94% over the baseline) and 72.40% on the 2-class task, establishing a new state-of-the-art on the 2-class task and competitive performance on the 4-class benchmark.
Audio language models are meant to answer questions from the audio itself, but an hour-long recording rarely fits in their context, and the answer often depends on a few seconds of it. Transcribing the recording first makes it fit but discards sounds and voices, while systems that return to the audio must be trained for this behavior. We propose ReListen, a training-free method in which a frozen audio language model finds its own evidence before answering. In a Jev-style scan, the model scores each 20-second window by the probability that it helps answer the question, and it then answers from the highest-scoring minute. This minute overlaps the annotated evidence far more often than transcript retrieval, for 88.5% of VoiceGiraffe questions. Without any change, the same scan also grounds events on LAT-Bench as well as a model trained for temporal grounding. Answering from it matches or exceeds hearing the whole recording with a small fraction of the audio, and on VoiceGiraffe it exceeds a transcript of the whole recording by 9.2 points. The selected evidence is also worth hearing rather than reading. Given the same excerpts, reading their transcript matches listening on speech content, but listening is clearly more accurate on questions about sounds and voices. Moreover, ReListen answers faster than every transcript-based system, and its memory stays constant even for recordings too long to hear whole on one GPU. Therefore, a frozen audio language model can search hour-long recordings by itself, without a transcript or additional training.