Guoqiang Hu, Lichao Zhang, Yizhou Peng, Yuhao Luo, Haifeng Hu, Sallen Sun, Will Yang, Eng Siong Chng
A full-duplex spoken dialogue model must yield when the user cuts in and hold when the user merely backchannels. Sequence-level RL with a timing reward (GRPO) is the usual recipe; we test its whole-rollout form on four open models and find two failures. A binary window reward gives zero advantage on exactly the windows that most need learning (a zero-gradient trap), and the one score per rollout is spread over listening-state decisions the reward is not written in terms of (credit leakage), so the model talks into silence or stops answering. Both come from sampling a reward that is a known function of a few binary control decisions. We propose \emph{window likelihood} (\WL): sample a continuation per window, but integrate in closed form over the stop decisions it visits and maximise the log-probability of acting correctly---a noisy-OR for yielding, a survival product for holding. On HumDial-FDBench, \WL\ gives the best yield/hold F1 on all six structurally different open models we post-train, exceeds Kyutai's released GRPO checkpoints (trained on different data), and adds no re-opening of the floor on any model.
Ziyun Chen, Yijun Wang, Chuanyi Zhang, Liang Yao, Yuexuan An, Xiang Gu, Wei Zhou, Fan Liu
Referring expression segmentation (RES) aims to localize targets at the pixel level guided by human instructions. However, current models predominantly follow a text-centric paradigm that treats discrete text as the primary interaction medium for these human instructions. This assumption is especially limiting in unmanned aerial vehicle (UAV) applications, where rapid response, hands-free operation, and natural communication are often critical. In this paper, we reconstruct UAV referring expression segmentation through an audio-centric perspective. Grounded in information theory, we demonstrate that raw audio provides a higher upper bound for extracting rich semantic features than transcribed text. To achieve this, we construct an audio-driven UAV referring expression segmentation benchmark named Aerial-AudioRES by augmenting VisDrone with synthesized speech, forming "audio-text-image-mask" quadruples. Specifically, we employ an ambiguity filtering mechanism to ensure referential uniqueness, and incorporate eight distinct timbres alongside 4 realistic noises to simulate authentic environments. To drive the paradigm shift from text-centric to audio-centric, we propose the end-to-end Audio-centric Referring Expression Segmentation Network (ARES-Net). Distinct from discrete text-visual matching, our approach learns to establish a direct soft alignment between continuous acoustic manifolds and multi-scale visual representations. Comprehensive experiments show that our method obtains a peak accuracy of 64.23% in oIoU and demonstrates robust generalization on a collected real-human test set. Furthermore, it significantly reduces total interaction latency to one-third of the text-based method. Our code is available at https://github.com/LemonGauss/ARES-Net.
Active learning (AL) reduces annotation by querying informative examples. Continual active learning (CAL) also cuts the training cost, the total number of gradient steps, by training one model across AL rounds with a fixed number of mini-batch steps per round. Pseudo-labels add supervision without annotation, but they increase training cost. In CAL, where every round trains on a fixed number of mini-batches, whether and how pseudo-labeled examples should share these mini-batches with labeled examples is therefore open. We provide the first formalization of this problem as two decisions after each query, namely (i) which unlabeled examples should enter training with a pseudo-label and (ii) what share of each mini-batch they should take. Our method, semi-supervised continual active learning (SSCAL), works with any acquisition function. It admits pseudo-labels based on distances from unlabeled examples to labeled neighbors, caps their share, and reweights the loss on labeled examples to keep their expected gradient contribution unchanged. On frozen embeddings, its pseudo-labeler updates incrementally without extra model forward passes. Across two audio datasets, two frozen encoders, six acquisition functions and ten seeds, SSCAL ranks first on average among four protocols with equal training steps, answering the question affirmatively.
Audio active learning (AL) trains a classifier at low annotation cost by labeling only an informative fraction of an unlabeled collection over successive rounds. Each round picks one of the checkpoints saved during training. Mainstream AL protocols do so with an assumed validation set, but in practice its labels also count toward the annotation cost, shrinking the advantage of an already small annotation budget. We propose geometric consistency, a label-free validation rule for checkpoint selection. Both classifiers, nearest-class-mean and k-nearest-neighbor, are built from the labeled set in a frozen pretrained embedding space. They then predict the unlabeled data to model its geometry. The rule scores each checkpoint by its agreement with these pseudo-labels and its cross-entropy on the labeled set. It works with any query strategy, has one hyperparameter, insensitive in practice, and adds only one batched forward pass over the unlabeled data per round. At matched label budgets across datasets, encoders, and query strategies, it attains the best mean rank in area under the labeling curve (AULC) among methods without extra labels and is on par with an out-of-budget validation set. To our knowledge, this is the first systematic study of label-free validation for checkpoint selection in audio AL.
A transcriber listening through one microphone also hears everyone near it. We train causal voice isolators: masking transformers of 140M, 243M and 374M parameters, 20 ms of latency, enrolling the caller from the call itself. Short evaluation windows hide two failures that real calls expose: forgetting the caller when another voice talks alone for a while, and damaging a caller who is alone. Long calls with turns, memory beyond the attention window, real headset crosstalk and drifting caller levels address both, but each trades one against the other. WavLM-Large's own weights, made causal and trained with the model, move the trade: M with them (555M parameters) lowers the error on our development sets from 11.44% to 10.55%; a student trained from scratch to imitate WavLM does not, and w2v-BERT 2.0 and Whisper's encoder, pre-trained on far more audio, help less or not at all. A lone caller loses most at the start of a recording: no training call opened with someone already talking. Training on windows cut from longer calls lowers the error on clean development calls from 3.92% to 2.96%; this is our release. Live, the release makes 16.13% word errors on 60 s AMI windows, against 23.66% for VoxLock, the strongest open isolator, and 10.19% on its own AMI windows, against 11.71%; it is level on Krisp's benchmark (5.74% against 5.63%), and a caller who is alone still loses more words than with VoxLock (5.58% against 2.92% on clean test calls). Distilled into 22.1M parameters without WavLM, it runs live on one CPU thread inside a LiveKit agent, at 33% of a core per call in 20 ms frames, with 10.66% error on our development sets (the release 10.49%). We open-source the release and this CPU model, with a LiveKit plugin, at https://huggingface.co/collections/Scicom-intl/voice-isolator.
Joseph An, Phillip Keung, Jiaqi Wang, Orevaoghene Ahia, Noah A. Smith
Audio language models process input audio into rich frame-level representations, but the standard approach to temporal localization generates timestamps as sequences of text tokens, which discards the frame-level representations in favor of autoregressive decoding. However, generating timestamps as tokens is slow and not parallelizable, and tends to hallucinate when producing timestamps outside the training distribution. We propose internal frame-level reuse, a method that trains audio LMs to reuse their own internal audio representations for temporal localization directly, bypassing token generation altogether. We introduce a lightweight prediction head trained via different frame-level objectives: a binary frame classifier and a novel inhomogeneous Poisson process (IHP) loss that models temporal event intensity. Across word localization, speaker diarization, and event localization tasks, our approach can achieve a >50x inference speedup over token-based generation and demonstrates robust length generalization, maintaining high accuracy on out-of-distribution audio durations where token-based models collapse completely. We find that reusing audio frame-level representations yields localization accuracies comparable to (and often better than) finetuned token-based baselines.
Geometric alignment scores based on Gram determinants provide a compact way to model higher-order consistency among modalities, yet how such scores respond to modality degradation is poorly understood. This paper asks whether the response of a multimodal geometric score is determined primarily by the magnitude of the perturbation-induced displacement. Using frozen cohorts from MSR-VTT (N=878) and DiDeMo (N=980), we apply controlled video blur and audio noise and analyze the response in the relational geometry on which the score is defined. Displacement magnitude explains at most 15% of the out-of-sample variance in the absolute response, and magnitude-matched pairs respond systematically differently, so scalar magnitude does not organize the response. The closed-form first-order expansion of the Gramian volume yields the Directional Geometric Response (DGR): the projection of the displacement onto the local volume gradient, which jointly captures the clean operating point, displacement magnitude, and displacement direction. The absolute first-order DGR term explains the observed response with out-of-sample R^2 of 0.838-0.969, matched-magnitude ranking accuracies of 0.864-0.963, and response-sign accuracies of 0.909-0.989, whereas the tested direction-free alternatives remain weak or unstable under the corresponding evaluation protocols. A pre-specified gain-normalization candidate, V/(g_V+eps), fails its predictability and clean-order gates. DGR uses the observed degraded-state displacement and is therefore an explanatory quantity, not a deployment-time predictor: geometric response depends on where the representation operates, how far degradation moves the relational geometry, and in which direction it moves.
Chain-of-thought TTS systems rely on language-model planners to translate persona and context into vocal delivery, but it is unclear where this intent is lost when the requested affect conflicts with lexical sentiment. We introduce a bilingual Japanese--English persona--line conflict benchmark in which each item pairs one line with congruent and conflicting personas and pre-declares the expected acoustic direction. We probe this direction in the planner-generated delivery phrase, the planner-written emotion vector, the QwenEmotion-generated vector, and the rendered waveform. Planner outputs are usually direction-correct, while downstream agreement can be substantially lower. Bypassing QwenEmotion provides only modest gains over the full planner-controlled Base. Supervising QwenEmotion raises the direction rate to 75% in Japanese and 84% in English. A learned text-to-latent interface reaches the highest English rate of 89% and the highest Japanese arousal rate of 88%. These results show that learned repair of the plan-to-synthesis interface can recover counter-textual delivery, although its effect differs across languages and acoustic axes.
Environmental noise identification is essential for noise control, yet conventional single-stage classifiers suffer from class confusion under imbalanced multiclass conditions. This paper proposes a lightweight hierarchical model integrating a Frequency-Time Adaptive Network (FTANet) with a two-stage classification mechanism. FTANet employs parallel asymmetric one-dimensional convolutions to capture temporal dynamics and frequency-band distributions, while adaptive fusion attention (AFA) enhances discriminative acoustic features. In the first stage, FTANet classifies sounds into five coarse-grained categories covering 25 fine-grained classes. In the second stage, samples are routed to the corresponding category-specific classifiers for fine-grained recognition. On a self-constructed dataset, FTANet achieves 95.45% coarse-grained accuracy with only 0.2004M parameters. On UrbanSound8K, it achieves 79.24% mean accuracy and 80.37% mean F1-score in ten-fold evaluation, outperforming compared convolutional models. Field tests on the Tianjin Online Acoustic Environment Monitoring Platform yield 87.94% coarse-grained accuracy and 97.21% conditional fine-grained accuracy for natural sounds correctly routed to C2. The complete hierarchical model achieves 87.18% overall accuracy and 87.40% weighted-average F1-score, demonstrating its feasibility for urban environmental noise monitoring.
oollim is a locally executable system for evaluating L2 English pronunciation accuracy, fluency, and prosody on a consumer computer. The system combines pretrained XLS-R, XEUS, and PhoneticXeus representations with reference-text features and ridge regressions trained on SpeechOcean762, the Korean Learners' L2 Speech Corpus (English), J-AESOP, and their combinations. On 305 independently held-out recordings assessed by three expert raters, several ridge configurations substantially outperform archival Microsoft Azure Pronunciation Assessment API data in both Spearman correlation and quadratic weighted kappa across all three constructs. These results support expert speech evaluation on consumer hardware using pretrained representations and construct-specific regressions, without neural-model fine-tuning or calibration on the private evaluation data.
Embodied assistants in domestic environments must infer what happened, where and when it occurred, and how to respond. To address this, we introduce RMS-AQA, a spatial audio question answering (SAQA) benchmark for real-world domestic environments. The benchmark features a two-stage question-answering (QA) format to comprehensively assess the ability of audio-language models (ALMs) to first ground audible sound events and subsequently perform complex spatio-temporal reasoning based on that grounding. To maximize acoustic realism, our dataset combines authentic real-world first-order Ambisonics (FOA) recordings with high-fidelity synthetic data generated using measured room impulse responses (RIRs). Furthermore, we provide a lightweight spatial plug-in that injects FOA-format data into frozen audio-language backbones. Experimental results reveal that the primary challenges stem from concurrent sources, far distance, and sim-to-real domain gap between RIR-synthesized and authentic recordings.
Pretrained image-generation models have been successfully repurposed for various downstream visual tasks. Whether their learned representations and generative priors can transfer beyond vision to other modalities remains largely unresolved, particularly for speech synthesis, which requires linguistic fidelity and preservation of speaker identity from a short reference. We present ImageTTS, which repurposes FLUX.2 for speech synthesis by encoding pseudo-RGB log-mel spectrograms with its frozen image VAE and adapting only its pretrained Diffusion Transformer (DiT) using conditional latent flow matching. This formulation disentangles two forms of vision-to-speech transfer: representation transfer through the image VAE and generative-prior transfer through DiT initialization. The frozen image VAE largely preserves intelligibility and speaker identity with 0.99 STOI and 0.95 speaker similarity, while image-pretrained DiT initialization accelerates convergence and improves linguistic accuracy, reducing WER from 4.2\% to 2.7\% at 100k matched updates relative to random initialization. The same framework supports both zero-shot and instruction-guided TTS and, when scaled to 100K hours of speech data, achieves 1.12\% WER on Seed-TTS test-en and a Style-ACC of 0.66 on CapSpeech, outperforming speech-native baselines trained on comparable data across both tasks. These results suggest that pretrained image models can transfer beyond architectural reuse, supplying useful latent representations and generative priors for intelligible and controllable speech synthesis.
Jincheng Wang, Li Li, Weixin Zhu, Yannan Wang, Tao Yu, Juan Liu, Ming Li
This paper presents S2SND-v2, a unified online and offline speaker diarization framework built on S2SND that improves acoustic feature extraction and the use of the resulting representations. For acoustic feature extraction, we introduce a speaker-adapted w2v-BERT 2.0 front end and compare full and pruned construction routes. To better use these features, overlap-aware activity-query masking improves speaker embedding extraction under overlap, while late-speaker training aligns unknown-speaker training with online inference conditions. Combining these improvements, the pruned system achieves competitive average diarization error rates of 14.52% online and 12.03% offline across eight test sets.
Aligned audio, performance MIDI, and musical scores provide valuable supervision for music analysis and performance modeling, yet datasets that jointly align all three modalities remain scarce and limited in scale and genre diversity. We introduce Gangqinpu, a multi-genre piano dataset containing 18,786 score-performance pairs and 1,007.1 hours of source audio, with automatic note-level alignments and derived beat/downbeat annotations. It spans diverse genres, including pop, classical, jazz, folk/traditional, rock, and blues. Our automated pipeline transcribes piano audio, aligns the resulting MIDI with musical scores, assigns genre labels, and filters out non-expressive performances with a score-informed binary classifier. We further construct joint splits with ASAP using conservative score-similarity grouping to reduce data leakage. We evaluate Gangqinpu on performance MIDI-to-score transcription, expressive performance rendering, and beat tracking. Downstream experiments demonstrate the value of Gangqinpu as complementary training data, with cross-dataset gains in beat tracking and primarily in-domain benefits for transcription and expressive performance rendering. The code and dataset are publicly available at https://github.com/wei-zeng98/gangqinpu.
Speech interaction systems have gradually evolved from traditional cascaded systems that use text as intermediate information to end-to-end speech language models that directly process speech representations. In some applications, the speech frontend runs on the user's device and sends speech representations to a remote server for inference, while the original audio remains local. However, these representations may still contain speaker identity information. If an attacker obtains them, it may use this information to imitate the user's voice, threatening their voice privacy. To investigate this risk, we propose \stia, a Speech Token Inversion Attack that generates speech with attacker-specified text in a victim's voice using a single short leaked representation, without access to the original audio or its transcript. Specifically, \stia converts discrete, continuous, or mixed speech representations into conditioning sequences and trains a conditional flow-matching speech generation model using pairs of different utterances from the same speaker in public speech datasets. The model combines speaker characteristics contained in the leaked representation with the specified text to generate speech in the victim's voice. Experiments show that, with a Moshi representation leak corresponding to 3 seconds of speech, \stia achieves an average CAM++ speaker cosine similarity of 0.78 and a target-text word error rate of 3.80\%. These results show that speech frontend representations reveal reusable speaker identity information and highlight the need to consider privacy risks when designing speech frontend representation methods.
Jihoon Jeong, Mirco Ravanelli, Zhepei Wang, Cem Subakan
Reasoning is increasingly important for large audio-language models (LALMs), yet remains relatively underexplored compared with text-only and vision-language models. We introduce Audio-R1, a reinforcement learning (RL) post-training framework that revisits audio self-supervised learning as a source of verifiable rewards for improving audio understanding and reasoning task performance. Audio-R1 uses answers automatically determined during the construction of self-supervised tasks to verify model responses and compute rewards. Experiments across multiple LALMs and diverse audio benchmarks show consistent performance gains, while ablations demonstrate the benefits of each self-supervised task and additional gains from joint training. While both supervised fine-tuning and Audio-R1 improve self-supervised task accuracy, Audio-R1 further improves performance on downstream tasks distinct from the training tasks. These results highlight the potential of RL post-training with self-supervised audio tasks to improve downstream performance and reasoning-trace quality.
Julia Wilkins, Wei-Ching Lin, Abinaya Kumar, Shabnam Ghaffarzadegan, Juan Pablo Bellow
Pretrained audio spectrogram transformers are widely used as backbones of machine listening systems for sound event classification. However, they are typically trained using spectrograms with a fixed Short-Time Fourier Transform (STFT) configuration, implicitly binding the model to the time-frequency resolution seen during pretraining. In practice, adapting these systems to new domains or computational constraints often requires different STFT frontend parameters. In this work, we conduct a controlled study of the sensitivity of the foundational Audio Spectrogram Transformer to variations in STFT window and hop size, showing that audio classification performance degrades substantially as frontend parameters deviate from pretraining settings. As a mitigation strategy, we introduce a lightweight conditioning mechanism using FiLM that enables a pretrained model to adapt to variable input configurations using limited additional training data. Our method generalizes well to unseen STFT configurations, improving model robustness without sacrificing performance in the original setting.
Kernel distances such as the Kernel Inception Distance (KID) and the Kernel Audio Distance (KAD) compare generated and reference samples in an embedding space, and model comparisons rest on their values. Each ships with one fixed kernel, and a reported value does not say which changes that kernel detects at the sample size used. We run each default as a permutation two-sample test on real embeddings of symbolic music, audio and images, comparing a sample with an independent sample changed by a mean shift, an isotropic contraction, or a rotation that keeps the mean and the covariance trace and is matched to the shift in transport distance. At 780 samples per side, both defaults rarely detect the rotations in five evaluation spaces, while they detect the shifts in every run. The median-bandwidth Gaussian kernel of KAD detects the contraction and regains power against rotations on image embeddings with an order of magnitude more samples. The cubic kernel of KID contains second- and third-moment information but misses the contraction in most runs, and our experiments do not establish a high-power rotation regime for it over the tested sample sizes. Narrower Gaussian bandwidths restore sensitivity in music and audio, an $\ell_1$ Laplace kernel does so for images, and aggregated tests combine both at nominal level. We recommend that a reported kernel distance state its sample budget and the changes its configuration detects at that budget.
Wendi Sang, Kai Li, Yifan Li, Jianqiang Huang, Xiaolin Hu
Streaming audio-visual target speaker extraction (AVTSE) is essential for latency-sensitive applications. However, current causal systems face two coupled gaps. First, standard benchmarks fix the number of speakers at two and ignore the near-/far-field interference, reverberation, and sudden noise of real scenes. Second, high-performance causal architectures are too heavy for edge deployment. Lightweight alternatives try to reclaim capacity by repeatedly invoking a small separator, which offsets the savings. To address the first gap, we release RealSSA, a realistic AVTSE benchmark consisting of RealSSA-Sim and RealSSA-Real. RealSSA-Sim provides controllable simulated scenes with 1--5 dynamic near-field interfering speakers; scenes with far-field interference additionally include 10--15 far-field interfering speakers. RealSSA-Real provides a held-out real-recorded evaluation set. To address the second gap, we propose Falcon, a causal AVTSE method. Its Separator performs multi-scale time-frequency separation in a single encoder--decoder pass. We train Falcon with an encoder-space speaker contrastive loss. This loss suppresses near-field leakage at zero inference cost and transfers across backbones. Falcon achieves state-of-the-art extraction quality on RealSSA-Sim, LRS2-2Mix, LRS3-2Mix, and VoxCeleb2-2Mix. Compared with the causal AV-TFGridNet baseline, it cuts parameters by 91.9%, computation by 20.2 times, and GPU inference latency by 11.1 times. Code and demo are available at https://cslikai.cn/Falcon/.
Yingzhuo Deng, Chong Li, Bingxin Liu, Yi Shu, Tianyu Peng, Can Xie, Siyao Song, Shuo Ren 等 9 位作者
Spoken Dialogue Models (SDMs) typically share a speech-token space for generation and synthesis, leading to long autoregressive sequences and costly inference. Lower-rate speech tokenizers shorten sequences but may introduce compression-quality trade-offs and costly replacement without breaking this coupling. To address this limitation, we propose CoSTA, a compatibility-preserving post-hoc framework that decouples generation and synthesis token spaces while retaining the original speech tokenizer and decoder. CoSTA extends the generation token space with compressed representations via mean-initialized vocabulary extension, pronunciation adaptation, and dialogue adaptation, followed by deterministic expansion back to the original speech-token space before synthesis. On bilingual speech-dialogue benchmarks, CoSTA achieves approximately $1.7\times$ expanded speech-token generation speed, lower full-response latency, and higher throughput while preserving response quality. Analyses show more sustained interaction under limited budgets and a practical trade-off among compression, decoding efficiency, and quality. Further experiments demonstrate cross-architecture transfer and complementarity to low-rate tokenization.
Bart van Merriënboer, Vincent Dumoulin, Jenny Hamer, Lauren Harrell, Andrea Burns, Tom Denton
Perch is a performant pre-trained model for bioacoustics. It was trained in supervised fashion, providing both off-the-shelf classification scores for thousands of vocalizing species as well as strong embeddings for transfer learning. In this new release, Perch 2.0, we expand from training exclusively on avian species to a large multi-taxa dataset. The model is trained with self-distillation using a prototype-learning classifier as well as a new source-prediction training criterion. Perch 2.0 obtains state-of-the-art performance on the BirdSet and BEANS benchmarks. It also outperforms specialized marine models on marine transfer learning tasks, despite having almost no marine training data. We present hypotheses as to why fine-grained species classification is a particularly robust pre-training task for bioacoustics.
Alvin Nahabwe, Sulaiman Kagumire, Denis Musinguzi, Bruno Beijuka, Jonah Mubuuke Kyagaba, Peter Nabende, Andrew Katumba, Joyce Nakatumba-Nabende
Automatic speech recognition (ASR) for African languages remains constrained by limited labeled data and the lack of systematic guidance on model selection, data scaling, and decoding strategies. Large pre-trained systems such as Whisper, XLS-R, MMS, and W2v-BERT have expanded access to ASR technology, but their comparative behavior in African low-resource contexts has not been studied in a unified and systematic way. In this work, we benchmark four state-of-the-art ASR models across 13 African languages, fine-tuning them on progressively larger subsets of transcribed data ranging from 1 to 400 hours. Beyond reporting error rates, we provide new insights into why models behave differently under varying conditions. We show that MMS and W2v-BERT are more data efficient in very low-resource regimes, XLS-R scales more effectively as additional data becomes available, and Whisper demonstrates advantages in mid-resource conditions. We also analyze where external language model decoding yields improvements and identify cases where it plateaus or introduces additional errors, depending on the alignment between acoustic and text resources. By highlighting the interaction between pre-training coverage, model architecture, dataset domain, and resource availability, this study offers practical and insights into the design of ASR systems for underrepresented languages.
Yunpeng Li, Kehang Han, Brian McWilliams, Zalan Borsos, Marco Tagliasacchi
We propose SpectroStream, a full-band multi-channel neural audio codec. Successor to the well-established SoundStream, SpectroStream extends its capability beyond 24 kHz monophonic audio and enables high-quality reconstruction of 48 kHz stereo music at bit rates of 4--16 kbps. This is accomplished with a new neural architecture that leverages audio representation in the time-frequency domain, which leads to better audio quality especially at higher sample rate. The model also uses a delayed-fusion strategy to handle multi-channel audio, which is crucial in balancing per-channel acoustic quality and cross-channel phase consistency.
Ben Williams, Bart van Merriënboer, Vincent Dumoulin, Jenny Hamer, Eleni Triantafillou, Abram B. Fleishman, Matthew McKown, Jill E. Munger 等 15 位作者
Machine learning has the potential to revolutionize passive acoustic monitoring (PAM) for ecological assessments. However, high annotation and compute costs limit the field's efficacy. Generalizable pretrained networks can overcome these costs, but high-quality pretraining requires vast annotated libraries, limiting its current applicability primarily to bird taxa. Here, we identify the optimum pretraining strategy for a data-deficient domain using coral reef bioacoustics. We assemble ReefSet, a large annotated library of reef sounds, though modest compared to bird libraries at 2% of the sample count. Through testing few-shot transfer learning performance, we observe that pretraining on bird audio provides notably superior generalizability compared to pretraining on ReefSet or unrelated audio alone. However, our key findings show that cross-domain mixing which leverages bird, reef and unrelated audio during pretraining maximizes reef generalizability. SurfPerch, our pretrained network, provides a strong foundation for automated analysis of marine PAM data with minimal annotation and compute costs.
Ye Jia, Michelle Tadmor Ramanovich, Quan Wang, Heiga Zen
We introduce CVSS, a massively multilingual-to-English speech-to-speech translation (S2ST) corpus, covering sentence-level parallel S2ST pairs from 21 languages into English. CVSS is derived from the Common Voice speech corpus and the CoVoST 2 speech-to-text translation (ST) corpus, by synthesizing the translation text from CoVoST 2 into speech using state-of-the-art TTS systems. Two versions of translation speeches are provided: 1) CVSS-C: All the translation speeches are in a single high-quality canonical voice; 2) CVSS-T: The translation speeches are in voices transferred from the corresponding source speeches. In addition, CVSS provides normalized translation text which matches the pronunciation in the translation speech. On each version of CVSS, we built baseline multilingual direct S2ST models and cascade S2ST models, verifying the effectiveness of the corpus. To build strong cascade S2ST baselines, we trained an ST model on CoVoST 2, which outperforms the previous state-of-the-art trained on the corpus without extra data by 5.8 BLEU. Nevertheless, the performance of the direct S2ST models approaches the strong cascade baselines when trained from scratch, and with only 0.1 or 0.7 BLEU difference on ASR transcribed translation when initialized from matching ST models.
Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, Marco Tagliasacchi
We present SoundStream, a novel neural audio codec that can efficiently compress speech, music and general audio at bitrates normally targeted by speech-tailored codecs. SoundStream relies on a model architecture composed by a fully convolutional encoder/decoder network and a residual vector quantizer, which are trained jointly end-to-end. Training leverages recent advances in text-to-speech and speech enhancement, which combine adversarial and reconstruction losses to allow the generation of high-quality audio content from quantized embeddings. By training with structured dropout applied to quantizer layers, a single model can operate across variable bitrates from 3kbps to 18kbps, with a negligible quality loss when compared with models trained at fixed bitrates. In addition, the model is amenable to a low latency implementation, which supports streamable inference and runs in real time on a smartphone CPU. In subjective evaluations using audio at 24kHz sampling rate, SoundStream at 3kbps outperforms Opus at 12kbps and approaches EVS at 9.6kbps. Moreover, we are able to perform joint compression and enhancement either at the encoder or at the decoder side with no additional latency, which we demonstrate through background noise suppression for speech.
Kevin Kilgour, Mauricio Zuluaga, Dominik Roblek, Matthew Sharifi
We propose the Fréchet Audio Distance (FAD), a novel, reference-free evaluation metric for music enhancement algorithms. We demonstrate how typical evaluation metrics for speech enhancement and blind source separation can fail to accurately measure the perceived effect of a wide variety of distortions. As an alternative, we propose adapting the Fréchet Inception Distance (FID) metric used to evaluate generative image models to the audio domain. FAD is validated using a wide variety of artificial distortions and is compared to the signal based metrics signal to distortion ratio (SDR), cosine distance and magnitude L2 distance. We show that, with a correlation coefficient of 0.52, FAD correlates more closely with human perception than either SDR, cosine distance or magnitude L2 distance, with correlation coefficients of 0.39, -0.15 and -0.01 respectively.