arXiv subjects

Andreas Stolcke

Publications and source records attributed to Andreas Stolcke.

At least 19 recordsLinked to original sources

arXiv2026

How Much Prosody Does Your ASR Encoder Already Know

End-of-turn (EoT) detection decides in real-time whether a speaker has finished. Traditional detectors rely on hand-built prosodic features (pitch, speaking rate, pauses), while modern detectors run on a neural ASR encoder that has processed the same audio. Because the encoder reads the signal those features are computed from, it may already hold much of the information they were meant to supply. We test this three ways, reading a streaming Zipformer encoder at four depths from raw input to the final layer. First, we measure how much prosody helps over the encoder alone. Second, we test whether prosody can be recovered from the encoder. Third, we ask whether any single feature still helps given the encoder. The three tests agree; prosody adds a large gain on raw audio, but that gain collapses once the signal passes through even the first encoder layer. The same pattern holds for two open-source encoders, wav2vec2 and Whisper, both of which are offline.

eess.AS↗

arXiv2026

The Importance of Speech Genre for End-of-Turn Detector Training

End-of-turn (EoT) detection has been studied mostly through features, models, and modalities, while the speech genre of the training data has drawn far less attention. We isolate this factor. Using one reproducible annotation pipeline, we build a matched pair of corpora, multi-turn human-human conversational and single-turn human-machine, and train the same streaming ConvLSTM EoT detector on each. Across four test sets the conversational model reaches up to +0.45 higher F1, driven by fewer false alarms on the mid-turn pauses and hesitations that are rare in single-turn corpora. This conversational-training effect is not specific to frame-level decoding. Under a matched complete-or-not test it also appears at the utterance level, across two encoders, but only on conversational test data. Testing only on single-turn speech, as most prior work does, therefore hides the benefit of conversational training data. Against open-source baselines, our 1.8M-parameter model attains the highest F1 on every test set.

eess.AS↗

arXiv2026

Towards Multilingual End-of-Turn Detection using Frozen Speech Encoders

A spoken dialogue system must judge when the user has yielded the floor. This end-of-turn (EoT) detector is normally trained and tuned afresh for each language, which is costly to scale to new languages. We instead build one on a frozen multilingual encoder, training only a small fixed-size head. Using five CALLHOME languages, we investigate three questions. First, how well a head trained on one language transfers zero-shot to the others. Second, how much target-language adaptation is required to recover full task performance. Third, whether a single fixed decision threshold holds across languages without per-language tuning. We find that the head transfers well, even between tonal and pitch-accent languages, and that a few hours of audio recover nearly all the loss, matching in-language training. A frozen multilingual encoder with one lightweight head is a practical, language-agnostic baseline for EoT detection.

eess.AS↗

arXiv2026

Pseudo-Labels as Rewards, Not Targets: Group-Relative Policy Optimization for Semi-Supervised ASR

Adapting ASR to a domain with only untranscribed audio typically relies on pseudo-labeling, a strong teacher transcribes the audio and a student is fine-tuned on it. Because supervised fine-tuning (SFT) treats every pseudo-label token as a target, the student inherits the teacher's errors. We show that using the same pseudo-labels as a reward reference for Group Relative Policy Optimization (GRPO) is more robust to label noise. In the banking domain, GRPO yields 29.3% to 56.2% relative WER improvements over SFT on Whisper-generated labels, and the advantage holds across four Whisper sizes, four teacher families, three domains and four languages. While SFT reproduces almost 100% of teacher errors verbatim, GRPO passes through only 27%. Rollout statistics explain why, erroneous pseudo-label words rarely appear in on-policy rollouts, so group-relative advantages stay aligned with the clean transcript, whereas SFT forces them into the policy's support and collapses the reward variance GRPO relies on.

cs.CL↗

arXiv2026

Residual Expert Routing for Text-Only Domain Adaptation of Factorized Transducer ASR

Fine-tuning the non-blank predictor of a factorized transducer on target-domain text improves recognition on that domain but degrades performance on the source domain. A single interpolation weight between base and adapted predictors cannot resolve this: the adapted predictor helps at some prefixes and hurts at others, often within one utterance. We propose conservative residual routing: the adapted predictor becomes a residual correction to the base, and a 16k-parameter gate decides at each prefix how much of it to admit. Closing the gate reproduces the base exactly. On a 40-hour conversational corpus, routing retains essentially all of the text-only gain on two held-out domains (19.89 to 16.74, 20.91 to 16.21 WER) while leaving source-domain WER unchanged (12.90 to 12.93); the adapted predictor alone costs 1.08 WER. Routing needs no per-domain tuning on target-domain audio. Analysis reveals the gate does not behave as a simple domain-level switch: it modulates intervention through the disagreement between the two predictors, suppressing it in-domain and amplifying it out-of-domain.

cs.CL↗

arXiv2026

ToolRACER: A Robust Agentic Conversation Emulation Resource for Agent Training and Evaluation

Task-oriented conversational agents remain fragile under real world conversation scenarios as they rarely follow a predictable script, especially when users exhibit non-cooperative behavior. Existing function-calling benchmarks often emphasize successful, cooperative interactions and underrepresent adversarial conversation trajectories, thereby limiting the training resources available for developing robust agents. We present ToolRACER, a synthetic data generation pipeline that coordinates user, assistant and tool emulation models to generate and validated multi-turn interactions between a user and an agent. Using ToolRACER, we construct ToolRACERBench a robust multi-turn conversation benchmark spanning six domains, ranging over 55 varied personas, generating a validated corpus of 5.6K conversation trajectories, with approximately 66\% of conversations containing failure-prone conversation scenarios. We inject adversarial behaviors, producing validated conversational interaction trajectories that capture realistic, robust scenarios. We evaluate models trained on ToolRACERBench against internal benchmarks, as well as on function calling benchmarks such as $τ^2$-bench, BFCLv3 and ACEBench to evaluate agentic accuracy and robustness. Models trained on ToolRACERBench improve end to end agentic accuracy across $τ^2$-bench and ACEBench, demonstrating significant gains when mixed with in-domain dataset in small language models for agent capability tasks.

cs.CL↗

arXiv2026

Abstract Compression of Conversational Audio Context for LLM-based ASR

Standard LLM-based speech recognition systems process utterances in isolation, limiting their ability to leverage conversational context. Here we study whether multimodal context from prior turns improves LLM-based ASR and how to represent that context efficiently. We find that, after supervised multi-turn training, conversational context mainly helps with the recognition of contextual entities. However, conditioning on raw context is expensive because the prior-turn audio embedding sequence grows rapidly with conversation length. To address this, we propose Abstract Compression, a learned latent abstraction of prior-turn audio that replaces each prior-turn audio sequence with a fixed number of latent tokens while retaining corresponding transcripts explicitly. On both in-domain and out-of-domain test sets, the compression-enabled model recovers part of the gains of raw-context conditioning with a smaller prior-turn audio footprint. We also provide targeted analyses of the compression setup and its tradeoffs.

cs.CL↗

arXiv2026

When Synthetic Speech Is All You Have: Better Call GRPO

LLM-based ASR adapted to regulated domains such as banking is bottlenecked by privacy: real speech is costly and legally constrained to collect, making synthetic text-to-speech (TTS) an attractive substitute. Yet synthetic speech stays acoustically mismatched with real recordings, and work on this gap has stayed within supervised fine-tuning (SFT). We instead turn to reinforcement learning, and show that Group Relative Policy Optimization (GRPO) extracts far more from the same synthetic speech than SFT. Synthetic-only adaptation of the model with GRPO, a critic-free method rewarding low-WER hypotheses, reduces WER by 40% relative to SFT (36.71%$\to$22.09%), and an SFT-then-GRPO combination pushes this further to 45%. We trace the gain to behavior rather than representation: GRPO reduces insertion errors by improving stopping calibration and speech-to-text alignment by better anchoring attention to audio, leaving early-layer representations intact. When synthetic speech is the main resource, reinforcement learning should be preferred over supervised fine-tuning.

cs.CL↗

arXiv2026

From Generative to Discriminative: Lightweight Adaptation of SpeechLLMs for Multimodal Emotion Recognition

SpeechLLMs can recognise emotion by generating a label token, but this approach relies on the LLM's existing output representations. We compare this generative approach with a discriminative alternative that uses a linear classification head on top of the LLM's hidden states. Because most multimodal emotion recognition studies evaluate with ground-truth transcripts, whereas deployed systems receive ASR output, we run every comparison with both, using a compact speechLLM (frozen Whisper-medium encoder, 3B LLM) on IEMOCAP and MELD. The classification head improves Macro F1 in 10 of 14 configurations. On IEMOCAP, it improves every configuration using ASR transcripts, reaching 76.97\% Macro F1, with a significant gain when using speech alone. On MELD, where ASR transcripts have a 32.38\% word error rate, it helps with speech alone but not with ASR transcripts. The benefit thus depends on both the dataset and the transcript source: in 4 of the 6 configurations that use text, the head's effect changes sign between ground-truth and ASR transcripts. Conclusions drawn with ground-truth transcripts therefore need not hold with ASR transcripts. Finally, comparing the learned class vectors with the LLM's vocabulary shows that, on IEMOCAP and MELD, their nearest tokens are mostly emotion-related words rather than explicit emotion labels.

cs.CL↗

arXiv2026

MADS: Multi-Agent Dialog Simulation for Robust Conversational Agents

Task-oriented conversational agents remain fragile under realistic conditions, especially when users exhibit non-cooperative behavior. Existing function-calling benchmarks underrepresent adversarial conversation trajectories, thereby limiting the training resources available for developing robust agents. We present MADS (Multi-Agent Dialog Simulation), a synthetic data generation pipeline that produces structured, diverse dialogue trajectories. We inject adversarial conversation, producing validated conversation trajectories that capture realistic, unhappy paths. Using MADS, we construct MADS-Bench: a benchmark spanning six domains, 55 personas, and 1095 conversations, with approximately 66% of conversations containing failure-prone conversation scenarios. We evaluate the robustness of the trained models using a two-stage training approach that combines supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR).

cs.CL↗

arXiv2026

Closing the Speech-Text Gap with Limited Audio for Effective Domain Adaptation in LLM-Based ASR

Conventional end-to-end automatic speech recognition (ASR) systems rely on paired speech-text data for domain adaptation. Recent LLM-based ASR architectures connect a speech encoder to a large language model via a projection module, enabling adaptation with text-only data. However, this introduces a modality gap, as the LLM is not exposed to the noisy representations produced by the speech projector. We investigate whether small amounts of speech can mitigate this mismatch. We compare three strategies: text-only adaptation, paired speech-text adaptation, and mixed batching (MB), which combines both. Experiments in in-domain and out-of-domain settings show that even limited speech consistently improves performance. Notably, MB using only 10% of the target-domain (less than 4 hours) speech achieves word error rates comparable to, or better than, conventional ASR fine-tuning with the full dataset, indicating that small amounts of speech provide a strong modality-alignment signal.

cs.CL↗

arXiv2026

Text-only adaptation in LLM-based ASR through text denoising

Adapting large language model (LLM)-based automatic speech recognition (ASR) systems to new domains using text-only data is a significant yet underexplored challenge. Standard fine-tuning of the LLM on the target domain text often disrupts the critical alignment between the speech and text modality learned by the projector, degrading performance. We introduce a novel text-only adaptation method that frames this process as a text denoising task. Our approach trains the LLM to recover clean transcripts from noisy inputs. This process effectively adapts the model to a target domain while preserving cross-modal alignment. Our solution is lightweight, requiring no architectural changes or additional parameters. Extensive evaluation on two datasets demonstrates up to 22.1% relative improvement, outperforming recent state-of-the-art text-only adaptation methods.

cs.SD↗

arXiv2026

Less can be More: What Aspects of Speech Drive End-of-Turn Detection

In conversational AI, detecting when a speaker has finished talking is crucial for natural turn taking. While recent work incorporates semantics, the relative contribution of different modalities remains unclear. We present a controlled ablation of acoustic, prosodic, and semantic signals for streaming end of turn detection using a lightweight trimodal classifier. Under identical training conditions, the acoustic prosodic combination achieves the best balance of accuracy and latency, achieving utterance F1 of 0.93 with 7.8% false alarms at 400ms median latency. Adding text increases premature detections without improving performance. Feature space analysis confirms that prosodic features have the strongest class separability, while text representations overlap substantially. These findings suggest that turn-taking is primarily conveyed through intonation and silence patterns rather than semantic completeness, enabling faster and more reliable systems without expensive text inference.

eess.AS↗

arXiv2026

LLMs Get Smarter from Targeted Synthetic Multilingual Data

Language-specific competency (LSC) is the phenomenon of a language model performing better or worse depending on the language of the prompt. In other words, a language model outputs different (and potentially incorrect) responses to the same semantic query when prompted in different languages. Prior work attributes this to an internal misalignment of semantic representation across languages. Currently, there are two main approaches to address LSC in the literature: (1) routing all queries through English, improving performance, but limiting language expressivity to English; or (2) training on language-balanced data, equalizing model performance across languages, but reducing overall performance. In this work, we take a data centric perspective and introduce HOTFIXR: Hardness Optimized Training data For Improving X-Lingual Reasoning. It is a data generation framework that uses models to probe and learn a student model's multilingual weaknesses, and generates data to mitigate them. HOTFIXR can generate multilingual synthetic training data that can improve multilingual performance. We evaluate on three in-distribution tasks, three out-of-distribution tasks, and four out-of-distribution languages. On average, HOTFIXR (1) improves in-distribution performance by 6.2%, (2) reduces catastrophic forgetting (induced by fine-tuning) on OOD tasks by 3.7%, and (3) on OOD languages by 7.1%. Overall, as many real-world applications requires multilingual LLMs, our work contributes to the efforts of making LLMs multilingually proficient. We will release code upon acceptance.

cs.CL↗

arXiv2026

How to Leverage Synthetic Speech for LLM-Based ASR Systems?

In regulated domains such as banking and healthcare, where privacy constraints make real speech costly to collect and retain, synthetic speech from modern text-to-speech (TTS) is an appealing alternative for training automatic speech recognition (ASR) without exposing sensitive customer recordings. Yet a persistent distributional gap between synthetic and real data limits how far it can replace genuine recordings. Prior work largely treats this gap as a black box to be engineered around, but in our work, we instead examine its origin directly by probing a SLAM-ASR architecture. Then, we localise where its LLM backbone separates real from synthetic speech and find the discriminative signal concentrated in the early-to-middle layers, where temporal and prosodic perturbations disrupt it most. We further show that representation-level separability, help, but does not directly predict downstream ASR gains. On the other hand, convolving synthetic audio with room impulse responses (RIRs) narrows the gap not by making synthetic speech sound cleaner or more natural, but by reproducing the acoustic irregularities of real recordings. Translating these findings into the training procedure, by adding a layer-selection module combined with RIR augmentation matches a fully real-data baseline using only 25% of the real speech (13.6h) and surpasses it at all higher proportions.

cs.CL↗

arXiv2026

Reducing Prompt Sensitivity in LLM-based Speech Recognition Through Learnable Projection

LLM-based automatic speech recognition (ASR), a well-established approach, connects speech foundation models to large language models (LLMs) through a speech-to-LLM projector, yielding promising results. A common design choice in these architectures is the use of a fixed, manually defined prompt during both training and inference. This setup not only enables applicability across a range of practical scenarios, but also helps maximize model performance. However, the impact of prompt design remains underexplored. This paper presents a comprehensive analysis of commonly used prompts across diverse datasets, showing that prompt choice significantly affects ASR performance and introduces instability, with no single prompt performing best across all cases. Inspired by the speech-to-LLM projector, we propose a prompt projector module, a simple, model-agnostic extension that learns to project prompt embeddings to more effective regions of the LLM input space, without modifying the underlying LLM-based ASR model. Experiments on four datasets show that the addition of a prompt projector consistently improves performance, reduces variability, and outperforms the best manually selected prompts.

eess.AS↗

arXiv2025

Spoken Conversational Agents with Large Language Models

Spoken conversational agents are converging toward voice-native LLMs. This tutorial distills the path from cascaded ASR/NLU to end-to-end, retrieval-and vision-grounded systems. We frame adaptation of text LLMs to audio, cross-modal alignment, and joint speech-text training; review datasets, metrics, and robustness across accents and compare design choices (cascaded vs. E2E, post-ASR correction, streaming). We link industrial assistants to current open-domain and task-oriented agents, highlight reproducible baselines, and outline open problems in privacy, safety, and evaluation. Attendees leave with practical recipes and a clear systems-level roadmap.

cs.CL↗

arXiv2025

Slot Filling as a Reasoning Task for SpeechLLMs

We propose integration of reasoning into speech large language models (speechLLMs) for the end-to-end slot-filling task. Inspired by the recent development of reasoning LLMs, we use a chain-of-thought framework to decompose the slot-filling task into multiple reasoning steps, create a reasoning dataset and apply the supervised fine-tuning strategy to a speechLLM. We distinguish between regular and reasoning speechLLMs and experiment with different types and sizes of LLMs as their text foundation models. We demonstrate performance improvements by introducing reasoning (intermediate) steps. However, we show that a reasoning textual LLM developed mainly for math, logic and coding domains might be inferior as a foundation model for a reasoning speechLLM. We further show that hybrid speechLLMs, built on a hybrid text foundation LLM and fine-tuned to preserve both direct and reasoning modes of operation, have better performance than those fine-tuned employing only one mode of operation.

cs.CL↗