Streaming speech recognizers output a word only after the speaker has finished it, typically 90-400 ms later. For a voice agent every millisecond counts: the sooner the words are outputted, the sooner the agent can start a search or plan its reply. We present Earlybird, a 29M-parameter streaming recognizer (16.8 MB at 4 bits). Earlybird learns from audio cut off at random points, including partway through a word, and is asked to predict every word that is due by then in full, plus a guess of the next two words. Its raw token probabilities give a confidence for every word including the one still being spoken. On clean audiobook speech (LibriSpeech test-clean), the median word is predicted correctly 90 ms before the speaker finishes it, while released streaming models get it 90 to 400 ms after it ends. Earlybird is the earliest on all six test sets we time, covering audiobooks, TED talks, public speech, earnings calls, and synthetic speech. On the same audiobooks, 69% of words are predicted correctly before they end, against 3-25% for the other models. For words still being spoken, confidence is calibrated, with an expected calibration error (ECE) of 3.7 to 5.5. On the eight English Open ASR Leaderboard sets, it is more accurate than Moonshine Tiny on every set and roughly ties Whistle, a model of the same size.
Which linguistic units minimise the bitrate required to encode speech content: phones, syllables, words, maybe larger units, or something in between? This question has implications for the development of speech and language systems and theories of human language processing. Using a transcribed corpus of 960 hours of English audiobooks, I compare the information rates of phones, syllables, words, multi-word combinations and intermediate tokenisations. While all units are equivalent under a perfect language model, it makes sense to compare these under limited modelling budgets; I use $n$-gram and transformer language models. The lowest bitrates are achieved using words, with slightly longer units sometimes performing best. The results show that, when modelling capacity is limited, units larger than syllables but smaller than long multi-word phrases provide the most efficient representation in terms of information rate.
We present Walrus, a unified 8B speech-and-audio foundation model handling speech recognition (including code-switching and multi-talker speech), speech-to-text and speech-to-speech translation, text-to-speech, emotion recognition, audio captioning, and general audio understanding. Our main contribution is not a new architecture but a data-centric training recipe; we use a standard decoder-only backbone initialized from Gemma 4, so that gains can be traced to the data. For speech generation, Walrus predicts low-bitrate discrete speech tokens, and a lightweight token-to-waveform decoder converts them into high-fidelity audio. The recipe is a three-stage data curriculum that moves from broad multilingual data to multi-talker, code-switching, and domain-specific data. At each stage, task-aware and dataset-aware weighting balances tasks with very different token budgets and protects core skills from forgetting. Trained on over 600M samples covering more than 20 tasks and 100 languages, Walrus leads on most multilingual, multi-talker, and code-switching ASR, speech translation, and audio understanding benchmarks, outperforming both open speech language models of comparable size and dedicated task-specific models. On most benchmarks, it even matches or surpasses the $4\times$ larger Qwen3-Omni's performance. Results suggest that data engineering is a primary driver of quality in multi-task speech and audio modeling.
Spoken Language Diarization (SLD) models provide frame-level language information of a speech recording. SLD modeling pipeline includes feature extraction from pre-trained speech models followed by clustering. Fine-tuning pre-trained speech models has shown performance gains for SLD with its magnitude being dataset dependent. Off-target Posterior Mass (OTPM), computed based on posterior language identification probabilities is proposed as a diagnostic metric to forecast performance gains from SLD fine-tuning. OTPM obtains positive correlation with probe gains, 0.56 on unseen synthetic bilingual data and 0.23 on real-world corpora. These results indicate promising potential of OTPM as a diagnostic metric for SLD.
We present GigaEdge Spench, an open-source SDK for streaming speech enhancement on CPU servers, mobile devices, and in browsers. Applications can build enhancement channels from neural denoising, normalization and dynamics blocks, sample-rate conversion, and acoustic echo cancellation where supported. These stages can be combined, while normalization can also run without a neural model. The release includes an 8 kHz INT8 denoiser and discriminative and hybrid 48 kHz FP32 models. A shared C++ core supports ONNX Runtime and TensorFlow Lite through native interfaces, Android and iOS SDKs, and WebAssembly integration. Independent stream contexts share model resources, and consecutive frames from one stream can be processed in a single inference call. We evaluate the released models on EARS-WHAM v2 and compare their quality and processing costs with RNNoise and DPDFNet2. All tested native and Web configurations process files faster than real time, although larger chunks require more time to collect input. We also measure how many streams a shared model can process on a CPU. Our contribution is a configurable deployment system and an experimental study of how model choice, chunk size, and execution environment affect its use. All reported SDK timing uses ONNX Runtime. Index Terms: speech enhancement, streaming inference, cross-platform deployment, temporal frame grouping, computational efficiency
When a user talks over a full-duplex dialogue model, it must decide whether the user is interrupting (stop) or backchannelling (carry on). Post-training can improve this discrimination or change how readily it stops. In our GRPO runs the second dominates: no configuration lowers both error rates on more than one model, and two released checkpoints discriminate no better than their bases. We trace this partly to how the reward is estimated. The probability of stopping inside a short window is a known function of a few stop probabilities the policy emits, yet GRPO estimates it by sampling. A group of rollouts that all agree gives no gradient, even when all fail; the push on the stop threshold is strongest where the model is unsure, whichever outcome is right; and whole-rollout credit reaches decisions the reward never judges. Confining the credit to the window removes only that leakage. Window likelihood (WL) samples the words but integrates over the stop decisions, so the probability of the correct outcome is computed exactly and its logarithm maximised. On HumDial-FDBench, across four open models, WL gives the best F1 and lowers both error rates by more than their confidence intervals, without talking into silence.
Data limitation is one of the most common issues in training machine learning classifiers for medical applications. Due to ethical concerns and data privacy, the number of people that can be recruited to such experiments is generally smaller than the number of participants contributing to non-healthcare datasets. Recent research showed that generative models can be used as an effective approach for data augmentation, which can ultimately help to train more robust classifiers sparse data domains. A number of studies proved that this data augmentation technique works for image and audio data sets. In this paper, we investigate the application of a similar approach to different types of speech and audio-based features extracted from interactions recorded with our automatic dementia detection system. Using two generative models we show how the generated synthesized samples can improve the performance of a DNN based classifier. The variational autoencoder increased the F-score of a four-way classifier distinguishing the typical patient groups seen in memory clinics from 58% to around 74%, a 16% improvement
This work proposes WaveGen, a fully end-to-end diffusion-based text-to-waveform framework based on Spectral Forcing, which jointly learns an internal spectral trajectory and the target waveform trajectory within a single model, enabling direct end-to-end text-to-waveform generation. Unlike recent TTS pipelines, WaveGen does not rely on separately trained or externally decoded intermediate representations, such as neural audio codecs, Mel-spectrogram vocoders, or VAE-based latent models. In addition, WaveGen performs internal text-speech alignment within a single model, eliminating the need for external alignment modules such as duration predictors. To preserve a truly end-to-end formulation, our framework further avoids dependence on self-supervised representations, such as BERT-based text embeddings and Wav2Vec 2.0-based semantic representations, as well as semantic distillation methods for training optimization. To achieve this, we carefully design Spectral Forcing, a model architecture with disentangled diffusion heads, and a training objective based on a multi-scale spectral $v$-loss. We further validate the scalability of end-to-end TTS models by scaling the model size from 0.1B to 0.6B parameters. Using a single architecture, open-source data, and one-stage training, without external modules such as duration predictor or REPA, WaveGen demonstrates fully end-to-end waveform generation achieving state-of-the-art performance on the Emilia training dataset. In particular, WaveGen achieves a WER of 1.47 and a SIM of 0.75 on the Seed-en benchmark. Source code and checkpoints are available at https://github.com/sh-lee-prml/WaveGen
We introduce AudioCodex, a fully causal autoencoder for live generative audio applications that compresses stereo audio into compact continuous latent sequences at ~10 Hz. It combines a Mean Flow magnitude autoencoder with latent hypersphere projection, bounded-noise augmentation, representation alignment, and adversarial phase recovery to learn representations that support both efficient downstream generation and high-quality reconstruction. Through extensive ablations, we show that AudioCodex outperforms the state-of-the-art in terms of generation-reconstruction trade-off, whilst remaining strictly causal in both encoding and decoding. This work paves the way towards truly real-time interactive and streaming generative audio applications.
We present BARS (Beat-Adaptive Rap Synthesis), a non-autoregressive rapping voice generation model based on Flow Matching, conditioned on a beat track (instrumental backing), lyrics, and a short reference voice for zero-shot timbre control. Unlike prior autoregressive rap synthesis, where rap pace and output duration are strictly determined by the beat, BARS enables independent control over the rap tempo, allowing flexible-length generation without explicit duration prediction or phoneme-level alignment. We encode lyrics using a pretrained ByT5 encoder with a lightweight Transformer adapter trained jointly with the model, without relying on speech-pretrained text encoders, enabling natural handling of non-standard orthography (e.g., slang, abbreviations) and numerals common in rap lyrics. To address alignment instabilities, such as token skipping or repetition, under strong beat conditioning, we propose Asymmetric Scaled RoPE for cross-attention, which introduces a positional structure to facilitate sequential lyric-audio alignment learning. Furthermore, the model supports partial re-generation via span masking, enabling lyric-level editing while preserving the surrounding vocal style. Experiments demonstrate that the proposed model achieves high lyric fidelity, natural rap flow that is rhythmically coherent with the beat, and robust tempo control across a wide range of rapping styles and speeds.
We propose a music inpainting technique combining similarity matching and a diffusion-based generative model to replace large segments of audio recordings with suitable alternatives. The proposed hybrid method, Similarity-Guided Diffusion Posterior Sampling (SimDPS), retrieves candidate replacement segments from the wider signal based on contextual similarity and uses them to guide the diffusion process toward coherent reconstructions. Being model-agnostic, the framework is compatible with any pre-trained diffusion model operating in either the waveform or latent domain. Objective evaluations demonstrate that SimDPS outperforms widely adopted RePaint and DPS methods across acoustic features and neural-embedding distance metrics. Subjective evaluations on multi-genre music datasets confirm that SimDPS enhances perceptual plausibility compared to previous diffusion-based methods and demonstrates robustness to poorly matched similarity candidates. These findings pave the way for improved music restoration and context-aware audio editing.
Many studies of animal vocalizations require clean recordings, but animals are usually recorded in the field, with background noise from weather, vehicles, and other animals. Animal vocalization denoising is challenged by the scarcity of clean training data and the extreme diversity of animal vocalizations. Classical spectral gates leave residual noise, especially without voice activity detection (VAD), while detector-free neural denoisers designed for bioacoustics remove less noise. We show that a general-purpose source separation foundation model, SAM-Audio, removes more noise than either of these types of methods, when used without training and prompted with time spans from a simple animal VAD. To turn this heavy method into a practical tool, we distill it into CallCleaner, a 19M-parameter denoiser that needs no detector and runs ~90x faster than real time on a laptop CPU. We find that SI-SDR has limited utility in our setting as it penalizes regenerated output and conflates noise removal with call fidelity, a trade-off our separate metrics reveal. CallCleaner leaves 11 dB less noise around the call than Biodenoising, the leading neural bioacoustic denoiser with public weights, and among fast, CPU-capable systems it pushes the Pareto frontier of noise removal vs. call fidelity, without requiring a detector. We release code and weights for CallCleaner, which enables fast, private, local denoising of single-caller recordings without specialized hardware.
Clinical speech is increasingly used for disease screening and clinical assessment, but directly distributing speech recordings poses privacy risks because they contain both speaker-identifying voiceprints and sensitive spoken content. Existing speech anonymisation approaches are primarily developed for general speech applications or task-specific clinical speech and overlook that different clinical tasks rely on distinct speech characteristics, necessitating task-adaptive anonymisation strategies to protect patient privacy while preserving clinical utility. Here we introduce CASA, a clinically adaptive speech anonymisation agent that transforms anonymisation from a fixed processing procedure into a task-guided optimisation process, dynamically tailoring voiceprint and content anonymisation to task-specific privacy and clinical utility requirements. Through recursive self-improvement (RSI), CASA iteratively evaluates and refines anonymisation strategies against task-specific privacy and clinical utility requirements, while learning from these outcomes to guide more efficient planning in subsequent clinical tasks. CASA is validated across 14 clinical tasks encompassing five languages, diverse elicitation paradigms, and mental health, cognitive and voice disorders. CASA preserved disease-related classification performance while meeting speaker-anonymisation requirements across all clinical tasks, versus only 8/14 for the strongest fixed method. Assessments of resistance to speaker re-identification by external attackers and identification and redaction of sensitive personal information in clinical speech further demonstrated CASA's dual protection of speaker identity and sensitive spoken content. Clinical expert assessments further supported the preservation of clinically relevant information and the acceptability of anonymised speech for controlled research sharing.
Diarization-conditioned target-speaker ASR (TS-ASR), in particular Diarization-conditioned Whisper (DiCoW), is an accurate and data-efficient approach to multi-talker transcription. It has, however, three structural limitations: the model transcribes each speaker independently, operates on short (30 s) windows, and optimizes its diarization and TS-ASR components separately. We address the first two. We integrate the diarization conditioning into a longform Parakeet backbone (DiCoP) and add cross-speaker attention, which lets the model attend across the speaker dimension so that speakers are no longer decoded in isolation, and an STNO mask refiner that re-estimates the diarization-derived conditioning. As conversational data with genuine multi-speaker interaction is scarce, we scale training with a pseudo-labeling pipeline that mines 5.4k hours of in-the-wild multi-speaker audio from YODAS2. Combining these with synthetic and real data yields state-of-the-art results across several meeting benchmarks. We open-source the resulting models, DiCoP and Pandemonium, and the pseudo-labeled data.
In the United States, most police departments communicate using police radio, which allows for officer-to-officer and officer-to-dispatch communication without depending on cellular network or WiFi availability. While some departments use encrypted digital radio, many continue to transmit in the clear, and are legally monitored by public citizens with `police scanners.' Because police radio is intended for internal police use and represents real-time communication, it may offer valuable insight into patterns of law enforcement activities without the potential bias and filtering introduced by a press release after the fact. In this study, we collect and examine over 14 million radio transmissions representing over 100,000 hours of speech by officers in 34 cities across the United States from July 2025 through July 2026; to our knowledge the largest scale analysis of Broadcast Police Communication (BPC) to date. We highlight both typical patterns of police communication and categorize anomalous activity that represents incidents including protests and major weather events. We present this dataset as a starting point for an array of police communication studies, conduct our own preliminary analysis, and propose several further areas of inquiry.
This paper presents the design and preliminary evaluation of KR A1, a humanoid-robot interaction prototype that integrates personalized voice synthesis, contextual dialogue processing, and emotion-aware response generation for potential mental-healthcare applications. The system architecture combines a locally hosted large language model (Qwen2.5-3B-Instruct, served via Ollama) for natural-language understanding and response generation, OpenVoice v2 for zero-shot speaker-adaptive speech synthesis, and a two-tier emotion-processing pipeline comprising a keyword heuristic and LLM-based inference. Face recognition (InsightFace) and vector memory retrieval (FAISS with sentence-transformer embeddings) enable user-personalized interaction. The experimental evaluation reported here is limited to the voice-cloning component: cloned voices were evaluated against reference recordings using eight acoustic similarity measures - mel-frequency cepstral coefficients, pitch, spectral centroid, spectral contrast, spectral rolloff, zero-crossing rate, formants, and speaker-embedding cosine similarity - supplemented by a same-speaker versus cross-speaker baseline comparison. All tests were conducted on a small, informal development sample; no statistical significance testing against a held-out baseline was performed. The emotion-recognition accuracy, end-to-end response latency, and therapeutic or clinical impact of the integrated system have not yet been formally evaluated and are identified as directions for future work. All testing was informal and self-conducted by the development team; no patients or external human participants were involved, and the system has not been evaluated or deployed in a clinical setting.
Speech retrieval finds relevant spoken documents given a spoken query. The spoken query alone may not reveal the retrieval goal, such as finding an answer or matching speaker identity. Instruction-aware speech retrieval expresses this goal through a natural-language text instruction. Large audio-language models (LALMs) effectively process speech and text, and our analysis shows that different retrieval requirements favor hidden states from different audio-encoder layers. We introduce IRIS-Embed, which adapts LALMs into embedding models through contrastive training and instruction-aware audio layer weighting. We construct about 900K training samples from real data and synthetic data, with instructions expressing requirements for one or more speech attributes. On INSPIRE, an instruction-aware speech retrieval benchmark, IRIS-Embed consistently outperforms the corresponding LALM baselines. The best model achieves 62.92% macro-average Recall@50, surpassing the strongest baseline at 45.15%. Controlled comparisons demonstrate instruction awareness and the benefits of audio layer weighting. On MAEB, a massive audio embedding benchmark, macro-average gains of 9.4-19.8 points over the corresponding baselines show generalization beyond speech retrieval.
We propose Stochastic Interpolant Prior for Speech (SIPS), a plug-and-play framework for speech enhancement and separation that augments pretrained predictors with a generative clean speech prior. Building on stochastic interpolants, we decompose the sampling dynamics into a predictor-induced drift and a score-based denoising component. The predictor guides sampling toward a task-consistent estimate, while a score model trained exclusively on clean speech promotes perceptual naturalness. Unlike prior hybrid approaches, which typically rely on architecture-specific conditioning and are tied to particular predictors or degradation settings, SIPS employs a degradation-agnostic prior that can be reused across predictors and additive degradation tasks without retraining. We evaluate SIPS on speech enhancement and separation using recent predictors such as SEMamba and FlexIO. Across both tasks, SIPS consistently improves non-intrusive speech quality metrics while maintaining competitive performance on reference-based distortion measures and word error rate. For speech enhancement, these improvements are further supported by a subjective listening test.
Automatic Speech Recognition (ASR) systems often struggle to accurately transcribe speech from Deaf and Hard of Hearing (DHH) speakers. Existing personalization approaches typically require substantial pre-recorded datasets and burden DHH users. We present EvolveCaptions, a collaborative, error-driven personalization system that connects caption correction with speaker-specific ASR adaptation. When transcribing a DHH speaker's speech, hearing partners repair transcription errors, and the system generates short, targeted clauses for the DHH speaker to record. These samples are used to incrementally fine-tune the model. In a study with 12 DHH and six hearing participants, EvolveCaptions reduced Word Error Rate (WER) across five scripted reading rounds, requiring an average of five minutes of additional speech recording. Participants found the interfaces usable but considered recording and correction effortful; acceptability depended on context and anticipated long-term benefit. These findings show the potential of collaborative, cumulative adaptation while exposing the labor and conditions needed to sustain it.
Large public datasets have driven recent progress in machine learning, from text and vision to speech and general audio. Music remains an outlier, as copyright blocks redistribution of training audio, while many modern tasks need individual stems rather than mixtures. Existing multitrack resources trade off realism, scale, and licensing, while synthetic alternatives rely on proprietary VSTs or naive open rendering that yields weak stems. We introduce SPDMX, a listening-guided rendering of public-domain PDMX scores into an openly-licensed stem corpus with aligned scores and an open render pipeline: $253{,}982$ songs, $490{,}159$ mono stems, and $5{,}469$ song-hours at $44.1$ kHz under CC BY 4.0, with paired MIDI and identifiers linking to PDMX metadata. $88{,}403$ songs ($34.8\%$; $2{,}734$ song-hours) provide multitrack supervision. To our knowledge, this is the largest openly-licensed synthetic multitrack subset by song-hours, enabling source separation, stem-conditioned synthesis, MIDI-to-audio, transcription, and joint symbolic-audio modeling. We release the corpus, metadata, and open pipeline for reuse on other symbolic collections. Demos and experimental notes are on the project page.