Speech Large Language Models

Speech foundation models and Speech LLMs have advanced speech understanding, yet deployment-oriented model selection is hindered by non-comparable evaluations caused by mismatched post-processing, and by training results that are hard to reproduce across data scales and pipelines. We present SURE, a unified experimentation framework that standardizes prediction formats, normalization, and scoring. SURE evaluates strong systems across paradigms, from conventional pipelines to Speech LLMs, on representative tasks under realistic acoustic and linguistic stressors. Beyond evaluation, SURE introduces an agent-assisted training conversion flow that maps paper and code into versioned, runnable training pipelines under a unified protocol on matched open-data subsets. Overall, SURE improves comparability and reproducibility for deployment-oriented evaluation.

Recent speech language models rely on encoders that are optimized separately from autoregressive models. Since these encoders are unaware of the downstream objectives, the extracted representations may not be optimal for downstream tasks. To address this limitation, we introduce a discrete latent variable model on mel spectrograms that jointly optimizes the encoder and the speech language model. Joint optimization not only brings improvements over codec-based and other mel-spectrogram-based baselines on zero-shot Text-to-Speech (TTS) and Speech-to-Text (STT) tasks, but also effectively alleviates common issues in autoregressive mel spectrogram modeling, such as prolonged silence generation and word omissions.

Data-aware post-training quantization (PTQ) minimizes a per-token reconstruction loss on a small calibration corpus, implicitly weighting positions by their empirical frequency. For \textbf{A}utomatic \textbf{S}peech \textbf{R}ecognition (ASR), this misaligns with tail-sensitive risk: names, numerals, and domain-specific words receive proportionally little calibration mass. We propose \textbf{Tail-Aware Reconstruction Quantization} (\TARQ), a label-free PTQ framework that shifts calibration toward the lexical tail via \textbf{\rareBAL}, a closed-form per-Linear-layer rule equalizing common/tail mass, paired with a metric-consistent residual correction. \TARQ\ requires no entity labels, no curated calibration set, no validation decoding, and no additional training. Across eight ASR backbones and six datasets at W4G128, \TARQ\ improves mean rare-\textbf{W}ord \textbf{E}rror \textbf{R}ate (rare-WER) without an aggregate-WER regression, achieves the lowest cross-corpus rare-WER swing among compared methods, and transfers to entity-rich benchmarks (ProfASR, ContextASR-Speech-En) without entity supervision.

Speech and audio systems operate in inherently non-stationary environments, yet continual learning (CL) research in this domain, especially in the foundation model era, remains fragmented that fail to account for the coupled, geometry-sensitive nature of acoustic representations. Modern speech foundation models operate over highly entangled, continuous representations that jointly encode linguistic, speaker, and paralinguistic factors within a shared latent space. CL is therefore fundamentally about preserving and evolving shared representation structure rather than retaining isolated task knowledge. In this work, we revisit CL for speech from a representation-centered perspective, and introduce a new taxonomy that organizes CL according to how underlying representation geometry evolves under non-stationary acoustic conditions. We further identify key mismatches between current CL assumptions and speech foundation model behavior, and finally outline a set of open challenges and future research directions.

Speech large language models (SLMs) are typically built from text large language model (TLM) checkpoints, yet they still suffer from a substantial modality gap. Prior work has mainly attempted to reduce this gap from the output side by making speech generation more text-like, but the gap remains. We argue that the key remaining bottleneck lies on the input side. We propose TextPro-SLM, an SLM that makes spoken input more closely resemble that of a prosody-aware text LLM. TextPro-SLM combines WhisperPro, a unified speech encoder that produces synchronized text tokens and prosody embeddings, with an LLM backbone trained to preserve the semantic capabilities of the original TLM while learning paralinguistic understanding. Experiments show that TextPro-SLM achieves the lowest modality gap among leading SLMs at both 3B and 7B scales, while also delivering strong overall performance on paralinguistic understanding tasks. These gains are achieved with only roughly 1,000 hours of LLM training audio, suggesting that reducing the modality gap from the input side is both effective and data-efficient.

Automatic Speech Recognition (ASR) and speaker diarization in Bangla remain challenging due to long form recordings, diverse acoustic conditions, and significant speaker variability. This work addresses these two core tasks in Bangla spoken language understanding by developing robust systems for long form ASR and speaker diarization. For ASR (Problem 1), we fine tune the tugstugi bengaliai regional asr whisper medium model on a custom-curated dataset of approximately 15,000 chunked and aligned Bangla audio segments, employing full weight training with extensive data augmentation including noise injection, reverb simulation, echo, clipping distortion, and pitch/time perturbation. For speaker diarization (Problem 2), we fine-tune the pyannote/segmentation-3.0 model using PyTorch Lightning on the competition annotated diarization dataset, swapping the fine-tuned segmentation backbone into the pyannote/speaker-diarization-community-1 pipeline while retaining the pretrained speaker embedding and clustering components. Our ASR system achieves a Word Error Rate (WER) of 0.2441, while our diarization system achieves a Diarization Error Rate (DER) of 0.2392, both evaluated on the test set, demonstrating notable improvements over the respective pretrained baselines. We describe our complete pipeline, including data preprocessing, text normalization, audio augmentation, training strategies, inference optimization, and post-processing for both tasks.

MiniMind-O is an open 0.1B-scale omni model built on the MiniMind language model. It accepts text, speech, and image inputs, and returns both text and streaming speech. The release includes model code, checkpoints, and the main Parquet training datasets for text-to-audio, image-to-text, and audio-to-audio training, making the complete interaction loop directly inspectable. The model uses a full MiniMind backbone as the Thinker and an independent four-layer Talker made from MiniMind blocks. Frozen SenseVoice-Small and SigLIP2 encoders provide speech and image features, which are mapped by lightweight MLP projectors and injected at modality-placeholder positions. The Talker reads a middle-layer Thinker state together with an autoregressive eight-layer Mimi-code buffer. Speaker control is handled by a dedicated speaker token, right-aligned reference codec prompts, and precomputed CAM++ speaker embeddings, so voice conditioning remains part of the audio-code context rather than a separate TTS module. With a 768-dimensional Talker, the dense and MoE variants reach average CERs of 0.0897 and 0.0900 in Thinker--Talker consistency evaluation, with overall voice-cloning similarities of 0.5995 and 0.5937. Beyond reporting a working system, the paper identifies three scale-critical design choices for small omni models: middle-layer semantic bridging, a released multimodal sequence format, and a parameter-efficient eight-codebook interface.

Existing voice deepfake detection and localization models rely heavily on representations extracted from speech foundation models (SFMs). However, downstream finetuning has now reached a state of diminishing returns. In this paper, we shift the focus to pretraining and propose a novel recipe that combines bottleneck masked embedding prediction with flow-matching based spectrogram reconstruction. The outcome, Alethia, is the first foundational audio encoder for various voice deepfake detection and localization tasks. We evaluate on $5$ different tasks with $56$ benchmark datasets, and note Alethia significantly outperforms state-of-the-art SFMs with superior robustness to real-world perturbations and zero-shot generalization to unseen domains (e.g., singing deepfakes). We also demonstrate the limitation of discrete targets in masked token prediction, and show the importance of continuous embedding prediction and generative pretraining for capturing deepfake artifacts.

Synthetic accented speech is a promising way to improve automatic speech recognition (ASR) when real accented recordings are scarce. We ask what makes such data useful for ASR fine-tuning: target-accent phoneme edits that expose the recognizer to accent-specific pronunciations, or random phoneme perturbations that act as augmentation in phoneme space. In a few-shot TTS pipeline, we compare LLM-generated accent edits with matched-rate random substitutions and oracle controls using ground-truth accented phonemes and prosody. Random substitutions recover much of the ASR gain: LLM target-accent edits improve over random by only a small margin, ground-truth phonemes stay close to the random baseline and nearly converge with it as the synthetic ASR fine-tuning set grows larger, and adding ground-truth prosody yields only a modest further gain. Mixing synthetic with real accented speech also stabilizes low-resource fine-tuning, but a fixed synthetic budget can later dilute the information in real data, showing that the real--synthetic ratio matters.

Real-time automatic speech recognition (ASR) systems face a fundamental trade-off between transcription accuracy and computational efficiency, particularly when deploying large-scale transformer models like Whisper. Existing streaming approaches either sacrifice accuracy through aggressive chunking or incur prohibitive memory costs through unbounded context accumulation. We present WhisperPipe, a novel streaming architecture that achieves bounded memory consumption while maintaining transcription quality through three key innovations a hybrid Voice Activity Detection (VAD) pipeline combining Silero VAD with energy-based filtering to reduce false activations by 34%, a dynamic buffering mechanism with overlapping context windows that prevents information loss at segment boundaries, and an adaptive processing strategy that balances latency and accuracy based on speech characteristics. Evaluated on 2.5 hours of diverse audio data, WhisperPipe demonstrates a median end-to-end latency of 89ms (90th percentile: 142ms) while consuming 48% less peak GPU memory and 80.9% lower average GPU utilization compared to baseline Whisper implementations. The system maintains stable memory usage over extended sessions, with zero growth rate across 150-minute continuous operation. Comparative analysis against related work shows that WhisperPipe achieves competitive accuracy (WER within 2% of offline Whisper) while operating at 3-5x lower latency than existing streaming solutions. The architecture's modular design enables deployment across resource-constrained environments, from edge devices to cloud infrastructure. Our results demonstrate that careful architectural design can reconcile the competing demands of real-time responsiveness and model sophistication in production ASR systems.

As part of the MediSpeech project, we aim to develop a system that transcribes and diarizes Dutch conversations between doctors and patients in real-time. In this research (in-progress) we explore ways of efficiently combining Whisper with speaker diarization (SD). After trying to prompt Whisper with text that contains speaker labels, we observed that it is able to insert labels into the transcription with promising accuracy. We continued this line of research by fine-tuning Whisper with speaker-labelled prompts to generate transcriptions in a format similar to that of Serialized Output Training (SOT). Fine-tuning Whisper yielded more consistent speaker IDs across the chunks of long-form audio and improved verbatim transcription. The study uncovered new challenges as Whisper's SD performance suffers because of mistakes that get propagated through prompts and inaccurate timestamps assigned to overlapping speech.

Simultaneous speech translation (SST) generates translations while receiving partial speech input. Recent advances show that large language models (LLMs) can substantially improve SST quality, but at the cost of high computational overhead. To reduce this cost, prior work reformulates SST as a multi-turn dialogue task, enabling full reuse of the LLM's key-value (KV) cache and eliminating redundant feature recomputation. However, this approach relies on supervised fine-tuning (SFT) data in dialogue form, for which few human annotations exist, and existing synthesis methods cannot guarantee data quality. In this work, we propose a Hierarchical Policy Optimization (HPO) approach that post-train models trained on imperfect SFT data. We introduce a hierarchical reward that balances translation quality and latency objectives. Experiments on English to Chinese/German/Japanese demonstrate improvements of over +7 COMET score and +1.25 MetricX score at a latency of 1.5 seconds. Comprehensive ablation studies further validate the effectiveness of different quality rewards, hierarchical reward formulations, and segmentation strategies. Code can be found here https://github.com/owaski/HPO

Hallucinations in Speech Large Language Models (SpeechLLMs) pose significant risks, yet existing detection methods typically rely on gold-standard outputs that are costly or impractical to obtain. Moreover, hallucination detection methods developed for text-based LLMs do not directly capture audio-specific signals. We investigate four attention-derived metrics: AUDIORATIO, AUDIOCONSISTENCY, AUDIOENTROPY, and TEXTENTROPY, designed to capture pathological attention patterns associated with hallucination, and train lightweight logistic regression classifiers on these features for efficient inference-time detection. Across automatic speech recognition and speech-to-text translation tasks, evaluations on Qwen-2-Audio and Voxtral-3B show that our approach outperforms uncertainty-based and prior attention-based baselines on in-domain data, achieving improvements of up to +0.23 PR-AUC, and generalises to out-of-domain ASR settings. We further find that strong performance can be achieved with approximately 100 attention heads, improving out-of-domain generalisation compared to using all heads. While effectiveness is model-dependent and task-specific training is required, our results demonstrate that attention patterns provide a valuable tool for hallucination detection in SpeechLLMs.

Speech Large Language Models route speech encoder representations into an LLM decoder that typically accounts for over 90% of total parameters. We study how much of this decoder capacity is actually needed for speech tasks. Across two LLM families and three scales (1-8B), we show that decoder redundancy is largely inherited from the pretrained LLM: text and speech inputs yield similar redundant blocks. We then measure excess capacity by pruning decoder layers and analysing post-pruning healing to increase robustness. Our findings show that 7-8B models retain good ASR performance with only 60% of decoder layers, and the same trend extends to smaller scales with reduced pruning tolerance. We then generalise to speech translation, and show that the same blocks of layers are redundant across speech encoders, tasks and languages, indicating that a more global redundancy structure exists, enabling a single pruned and multi-tasks SpeechLLM backbone to be deployed.

Recent advances in automatic speech recognition (ASR) and speech enhancement have strengthened the common belief that cleaner audio should lead to more accurate transcription. In this work, we examine whether this assumption holds for modern zero-shot ASR systems. We conduct a structured empirical study of SAM-Audio as a preprocessing step for zero-shot transcription with OpenAI Whisper. Five Whisper variants are evaluated on noisy Bengali and English speech datasets. On the English dataset, SAM-Audio increases the average PSNR from 32.28 dB to 35.99 dB and achieves higher PSNR for 71.84% of the utterances. However, WER and CER increase in every evaluated model-dataset configuration. On the Bengali dataset, Whisper large-v3 WER increases from 65.83% to 77.35%, while CER increases from 24.13% to 34.74%. On the English dataset, Whisper base WER increases from 10.53% to 21.66%, while CER increases from 4.48% to 12.50%. Utterance-level analysis further shows that the degradation affects a substantial portion of the evaluated samples, although its severity varies across Whisper variants. These findings demonstrate that improved signal-level quality does not necessarily lead to better zero-shot ASR performance and that denoising can reduce recognition accuracy.

Scaling Multimodal Large Language Models (MLLMs) to long-form speech is bottlenecked by the explosive growth of input tokens. Existing speech-language models project high-frame-rate acoustic features directly into the LLM input space, making long-context processing computationally prohibitive. Unlike images or videos, speech lacks spatial redundancy, making extreme token compression particularly challenging. To address this limitation, we propose FastSLM, a token-efficient architecture featuring the Hierarchical Temporal Abstractor (HTA), which progressively distills acoustic features across multiple temporal scales. HTA achieves an extreme compression rate of 1.67 tokens per second (97% reduction) while preserving essential linguistic information for downstream speech-language understanding. Experimental results demonstrate that FastSLM achieves competitive performance across diverse speech-language tasks while requiring substantially fewer speech tokens and FLOPs than existing speech-language models. The source code and model checkpoints are available at https://github.com/Lee-junseok1025/FastSLM.

Deepfakes have emerged as a widespread and rapidly escalating concern in generative AI, spanning images, audio, and videos. Among these, audio deepfakes are particularly alarming due to the growing accessibility of high-quality voice synthesis tools and the ease with which synthetic speech can be distributed through social media and robocalls. Consequently, detecting audio deepfakes is critical for combating the misuse of AI-generated speech. However, real-world audio is often affected by corruptions such as noise, audio modification, and compression, which can significantly degrade detection performance. In this work, we systematically evaluate the robustness of 10 audio deepfake detection models against 18 common corruption types, grouped into three categories: noise perturbation, audio modification, and compression. Using both traditional deep learning models and state-of-the-art speech foundation models, our study yields four key insights. (1) Most models are robust to noise but remain vulnerable to audio modifications and compression, especially neural codecs. (2) Speech foundation models consistently outperform traditional models across most corruption scenarios, likely due to large-scale pre-training on diverse audio datasets. (3) Increasing model size improves robustness, although the gains diminish as models become larger. (4) Robustness to unseen corruptions can be improved through targeted data augmentation during training or speech enhancement at inference time. These findings highlight the importance of evaluating audio deepfake detectors under diverse real-world corruptions and developing more robust detection frameworks for practical deployment. We further advocate that future research on deepfake detection across all media should account for the diverse and unpredictable distortions encountered in real-world environments.

The rapid emergence of new entities -- driven by cultural shifts, evolving trends, and personalized user data -- poses a significant challenge for existing Speech Large Language Models (Speech LLMs). While these models excel at general conversational tasks, their static training knowledge limits their ability to recognize domain-specific terms such as contact names, playlists, or technical jargon. Existing solutions primarily rely on prompting, which suffers from poor scalability: as the entity list grows, prompting encounters context window limitations, increased inference latency, and the "lost-in-the-middle" phenomenon. An alternative approach, Generative Error Correction (GEC), attempts to rewrite transcripts via post-processing but frequently suffers from "over-correction", introducing hallucinations of entities that were never spoken. In this work, we introduce LOGIC (Logit-Space Integration for Contextual Biasing), an efficient and robust framework that operates directly in the decoding layer. Unlike prompting, LOGIC decouples context injection from input processing, ensuring constant-time complexity relative to prompt length. Extensive experiments using the Phi-4-MM model across 11 multilingual locales demonstrate that LOGIC achieves an average 9% relative reduction in Entity WER with a negligible 0.30% increase in False Alarm Rate.

We present BaltiVoice, a 16.8-hour read-speech corpus for Balti (ISO 639-3: bft), a Tibetic language spoken in Gilgit-Baltistan, Pakistan, with no prior publicly available ASR resources. The corpus contains 10,060 validated utterances in native Nastaliq script, derived from Mozilla Common Voice recordings. Fine-tuning OpenAI Whisper-small yields a Word Error Rate (WER) of 24.78% and a Character Error Rate (CER) of 8.30% after training for 5 epochs (3,000 steps) on the 538-utterance speaker-disjoint validation set, down from a zero-shot baseline of 159.19% WER and 152.52% CER. A Whisper-base fine-tuned on the same data achieves 44.54% WER and 15.61% CER, confirming that model capacity matters for this low-resource setting. The dataset, fine-tuned model, and a live transcription demo are publicly available on HuggingFace.