Hagen Soltau

Yu Zhang, Wei Han, James Qin et al. · meta-ai

We introduce the Universal Speech Model (USM), a single large model that performs automatic speech recognition (ASR) across 100+ languages. This is achieved by pre-training the encoder of the model on a large unlabeled multilingual dataset of 12 million (M) hours spanning over 300 languages, and fine-tuning on a smaller labeled dataset. We use multilingual pre-training with random-projection quantization and speech-text modality matching to achieve state-of-the-art performance on downstream multilingual ASR and speech-to-text translation tasks. We also demonstrate that despite using a labeled training set 1/7-th the size of that used for the Whisper model, our model exhibits comparable or better performance on both in-domain and out-of-domain speech recognition tasks across many languages.

Mingqiu Wang, Wei Han, Izhak Shafran et al. · deepmind

We present a joint Speech and Language Model (SLM), a multitask, multilingual, and dual-modal model that takes advantage of pretrained foundational speech and language models. SLM freezes the pretrained foundation models to maximally preserves their capabilities, and only trains a simple adapter with just 1\% (156M) of the foundation models' parameters. This adaptation not only leads SLM to achieve strong performance on conventional tasks such as speech recognition (ASR) and speech translation (AST), but also introduces the novel capability of zero-shot instruction-following for more diverse tasks: given a speech input and a text instruction, SLM is able to perform unseen generation tasks including contextual biasing ASR using real-time context, dialog generation, speech continuation, and question answering, etc. Our approach demonstrates that the representational gap between pretrained speech and language models might be narrower than one would expect, and can be bridged by a simple adaptation mechanism. As a result, SLM is not only efficient to train, but also inherits strong capabilities already acquired in foundation models of different modalities.

Jeffrey Zhao, Yuan Cao, Raghav Gupta et al. · deepmind

We propose AnyTOD, an end-to-end, zero-shot task-oriented dialog (TOD) system capable of handling unseen tasks without task-specific training. We view TOD as a program executed by a language model (LM), where program logic and ontology is provided by a designer as a schema. To enable generalization to unseen schemas and programs without prior training, AnyTOD adopts a neuro-symbolic approach. A neural LM keeps track of events occurring during a conversation and a symbolic program implementing the dialog policy is executed to recommend next actions AnyTOD should take. This approach drastically reduces data annotation and model training requirements, addressing the enduring challenge of rapidly adapting a TOD system to unseen tasks and domains. We demonstrate state-of-the-art results on STAR, ABCD and SGD benchmarks. We also demonstrate strong zero-shot transfer ability in low-resource settings, such as zero-shot on MultiWOZ. In addition, we release STARv2, an updated version of the STAR dataset with richer annotations, for benchmarking zero-shot end-to-end TOD models.

Mingqiu Wang, Izhak Shafran, Hagen Soltau et al. · deepmind

Large Language Models (LLMs) have been applied in the speech domain, often incurring a performance drop due to misaligned between speech and language representations. To bridge this gap, we propose a joint speech and language model (SLM) using a Speech2Text adapter, which maps speech into text token embedding space without speech information loss. Additionally, using a CTC-based blank-filtering, we can reduce the speech sequence length to that of text. In speech MultiWoz dataset (DSTC11 challenge), SLM largely improves the dialog state tracking (DST) performance (24.7% to 28.4% accuracy). Further to address errors on rare entities, we augment SLM with a Speech2Entity retriever, which uses speech to retrieve relevant entities, and then adds them to the original SLM input as a prefix. With this retrieval-augmented SLM (ReSLM), the DST performance jumps to 34.6% accuracy. Moreover, augmenting the ASR task with the dialog understanding task improves the ASR performance from 9.4% to 8.5% WER.

Dian Yu, Mingqiu Wang, Yuan Cao et al. · deepmind

Carefully-designed schemas describing how to collect and annotate dialog corpora are a prerequisite towards building task-oriented dialog systems. In practical applications, manually designing schemas can be error-prone, laborious, iterative, and slow, especially when the schema is complicated. To alleviate this expensive and time consuming process, we propose an unsupervised approach for slot schema induction from unlabeled dialog corpora. Leveraging in-domain language models and unsupervised parsing structures, our data-driven approach extracts candidate slots without constraints, followed by coarse-to-fine clustering to induce slot types. We compare our method against several strong supervised baselines, and show significant performance improvement in slot schema induction on MultiWoz and SGD datasets. We also demonstrate the effectiveness of induced schemas on downstream applications including dialog state tracking and response generation.

Hagen Soltau, Izhak Shafran, Mingqiu Wang et al. · deepmind

Most research on task oriented dialog modeling is based on written text input. However, users interact with practical dialog systems often using speech as input. Typically, systems convert speech into text using an Automatic Speech Recognition (ASR) system, introducing errors. Furthermore, these systems do not address the differences in written and spoken language. The research on this topic is stymied by the lack of a public corpus. Motivated by these considerations, our goal in hosting the speech-aware dialog state tracking challenge was to create a public corpus or task which can be used to investigate the performance gap between the written and spoken forms of input, develop models that could alleviate this gap, and establish whether Text-to-Speech-based (TTS) systems is a reasonable surrogate to the more-labor intensive human data collection. We created three spoken versions of the popular written-domain MultiWoz task -- (a) TTS-Verbatim: written user inputs were converted into speech waveforms using a TTS system, (b) Human-Verbatim: humans spoke the user inputs verbatim, and (c) Human-paraphrased: humans paraphrased the user inputs. Additionally, we provided different forms of ASR output to encourage wider participation from teams that may not have access to state-of-the-art ASR systems. These included ASR transcripts, word time stamps, and latent representations of the audio (audio encoder outputs). In this paper, we describe the corpus, report results from participating teams, provide preliminary analyses of their results, and summarize the current state-of-the-art in this domain.

Dian Yu, Mingqiu Wang, Yuan Cao et al. · deepmind

Knowledge (including structured knowledge such as schema and ontology, and unstructured knowledge such as web corpus) is a critical part of dialog understanding, especially for unseen tasks and domains. Traditionally, such domain-specific knowledge is encoded implicitly into model parameters for the execution of downstream tasks, which makes training inefficient. In addition, such models are not easily transferable to new tasks with different schemas. In this work, we propose to perform dialog state tracking grounded on knowledge encoded externally. We query relevant knowledge of various forms based on the dialog context where such information can ground the prediction of dialog states. We demonstrate superior performance of our proposed method over strong baselines, especially in the few-shot learning setting.

Hagen Soltau, Izhak Shafran, Alex Ottenwess et al.

We propose a Perceiver-based sequence classifier to detect abnormalities in speech reflective of several neurological disorders. We combine this classifier with a Universal Speech Model (USM) that is trained (unsupervised) on 12 million hours of diverse audio recordings. Our model compresses long sequences into a small set of class-specific latent representations and a factorized projection is used to predict different attributes of the disordered input speech. The benefit of our approach is that it allows us to model different regions of the input for different classes and is at the same time data efficient. We evaluated the proposed model extensively on a curated corpus from the Mayo Clinic. Our model outperforms standard transformer (80.9%) and perceiver (81.8%) models and achieves an average accuracy of 83.1%. With limited task-specific data, we find that pretraining is important and surprisingly pretraining with the unrelated automatic speech recognition (ASR) task is also beneficial. Encodings from the middle layers provide a mix of both acoustic and phonetic information and achieve best prediction results compared to just using the final layer encodings (83.1% vs. 79.6%). The results are promising and with further refinements may help clinicians detect speech abnormalities without needing access to highly specialized speech-language pathologists.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Ying Ma, Owen Burns, Mingqiu Wang et al.

Reinforcement learning (RL) is an effective method of finding reasoning pathways in incomplete knowledge graphs (KGs). To overcome the challenges of a large action space, a self-supervised pre-training method is proposed to warm up the policy network before the RL training stage. To alleviate the distributional mismatch issue in general self-supervised RL (SSRL), in our supervised learning (SL) stage, the agent selects actions based on the policy network and learns from generated labels; this self-generation of labels is the intuition behind the name self-supervised. With this training framework, the information density of our SL objective is increased and the agent is prevented from getting stuck with the early rewarded paths. Our self-supervised RL (SSRL) method improves the performance of RL by pairing it with the wide coverage achieved by SL during pretraining, since the breadth of the SL objective makes it infeasible to train an agent with that alone. We show that our SSRL model meets or exceeds current state-of-the-art results on all Hits@k and mean reciprocal rank (MRR) metrics on four large benchmark KG datasets. This SSRL method can be used as a plug-in for any RL architecture for a KGR task. We adopt two RL architectures, i.e., MINERVA and MultiHopKG as our baseline RL models and experimentally show that our SSRL model consistently outperforms both baselines on all of these four KG reasoning tasks. Full code for the paper available at https://github.com/owenonline/Knowledge-Graph-Reasoning-with-Self-supervised-Reinforcement-Learning.

Mingqiu Wang, Izhak Shafran, Hagen Soltau et al. · deepmind

We recently developed SLM, a joint speech and language model, which fuses a pretrained foundational speech model and a large language model (LLM), while preserving the in-context learning capability intrinsic to the pretrained LLM. In this paper, we apply SLM to speech dialog applications where the dialog states are inferred directly from the audio signal. Task-oriented dialogs often contain domain-specific entities, i.e., restaurants, hotels, train stations, and city names, which are difficult to recognize, however, critical for the downstream applications. Inspired by the RAG (retrieval-augmented generation) paradigm, we propose a retrieval augmented SLM (ReSLM) that overcomes this weakness. We first train a speech retriever to retrieve text entities mentioned in the audio. The retrieved entities are then added as text inputs to the underlying SLM to bias model predictions. We evaluated ReSLM on speech MultiWoz task (DSTC-11 challenge), and found that this retrieval augmentation boosts model performance, achieving joint goal accuracy (38.6% vs 32.7%), slot error rate (20.6% vs 24.8%) and ASR word error rate (5.5% vs 6.7%). While demonstrated on dialog state tracking, our approach is broadly applicable to other speech tasks requiring contextual information or domain-specific entities, such as contextual ASR with biasing capability.

Austin Stone, Hagen Soltau, Robert Geirhos et al.

Visual imagery does not consist of solitary objects, but instead reflects the composition of a multitude of fluid concepts. While there have been great advances in visual representation learning, such advances have focused on building better representations for a small number of discrete objects bereft of an understanding of how these objects are interacting. One can observe this limitation in representations learned through captions or contrastive learning -- where the learned model treats an image essentially as a bag of words. Several works have attempted to address this limitation through the development of bespoke learned architectures to directly address the shortcomings in compositional learning. In this work, we focus on simple, and scalable approaches. In particular, we demonstrate that by substantially improving weakly labeled data, i.e. captions, we can vastly improve the performance of standard contrastive learning approaches. Previous CLIP models achieved near chance rate on challenging tasks probing compositional learning. However, our simple approach boosts performance of CLIP substantially and surpasses all bespoke architectures. Furthermore, we showcase our results on a relatively new captioning benchmark derived from DOCCI. We demonstrate through a series of ablations that a standard CLIP model trained with enhanced data may demonstrate impressive performance on image retrieval tasks.

Hagen Soltau, Izhak Shafran, Mingqiu Wang et al.

Popular solutions to Named Entity Recognition (NER) include conditional random fields, sequence-to-sequence models, or utilizing the question-answering framework. However, they are not suitable for nested and overlapping spans with large ontologies and for predicting the position of the entities. To fill this gap, we introduce a new model for NER task -- an RNN transducer (RNN-T). These models are trained using paired input and output sequences without explicitly specifying the alignment between them, similar to other seq-to-seq models. RNN-T models learn the alignment using a loss function that sums over all alignments. In NER tasks, however, the alignment between words and target labels are available from the human annotations. We propose a fixed alignment RNN-T model that utilizes the given alignment, while preserving the benefits of RNN-Ts such as modeling output dependencies. As a more general case, we also propose a constrained alignment model where users can specify a relaxation of the given input alignment and the model will learn an alignment within the given constraints. In other words, we propose a family of seq-to-seq models which can leverage alignments between input and target sequences when available. Through empirical experiments on a challenging real-world medical NER task with multiple nested ontologies, we demonstrate that our fixed alignment model outperforms the standard RNN-T model, improving F1-score from 0.70 to 0.74.

Mingqiu Wang, Hagen Soltau, Laurent El Shafey et al.

Confidence estimate is an often requested feature in applications such as medical transcription where errors can impact patient care and the confidence estimate could be used to alert medical professionals to verify potential errors in recognition. In this paper, we present a lightweight neural confidence model tailored for Automatic Speech Recognition (ASR) system with Recurrent Neural Network Transducers (RNN-T). Compared to other existing approaches, our model utilizes: (a) the time information associated with recognized words, which reduces the computational complexity, and (b) a simple and elegant trick for mapping between sub-word and word sequences. The mapping addresses the non-unique tokenization and token deletion problems while amplifying differences between confusable words. Through extensive empirical evaluations on two different long-form test sets, we demonstrate that the model achieves a performance of 0.4 Normalized Cross Entropy (NCE) and 0.05 Expected Calibration Error (ECE). It is robust across different ASR configurations, including target types (graphemes vs. morphemes), traffic conditions (streaming vs. non-streaming), and encoder types. We further discuss the importance of evaluation metrics to reflect practical applications and highlight the need for further work in improving Area Under the Curve (AUC) for Negative Precision Rate (NPV) and True Negative Rate (TNR).

Hagen Soltau, Mingqiu Wang, Izhak Shafran et al.

In this paper, we describe novel components for extracting clinically relevant information from medical conversations which will be available as Google APIs. We describe a transformer-based Recurrent Neural Network Transducer (RNN-T) model tailored for long-form audio, which can produce rich transcriptions including speaker segmentation, speaker role labeling, punctuation and capitalization. On a representative test set, we compare performance of RNN-T models with different encoders, units and streaming constraints. Our transformer-based streaming model performs at about 20% WER on the ASR task, 6% WDER on the diarization task, 43% SER on periods, 52% SER on commas, 43% SER on question marks and 30% SER on capitalization. Our recognizer is paired with a confidence model that utilizes both acoustic and lexical features from the recognizer. The model performs at about 0.37 NCE. Finally, we describe a RNN-T based tagging model. The performance of the model depends on the ontologies, with F-scores of 0.90 for medications, 0.76 for symptoms, 0.75 for conditions, 0.76 for diagnosis, and 0.61 for treatments. While there is still room for improvement, our results suggest that these models are sufficiently accurate for practical applications.

Izhak Shafran, Nan Du, Linh Tran et al.

There is a growing interest in creating tools to assist in clinical note generation using the audio of provider-patient encounters. Motivated by this goal and with the help of providers and medical scribes, we developed an annotation scheme to extract relevant clinical concepts. We used this annotation scheme to label a corpus of about 6k clinical encounters. This was used to train a state-of-the-art tagging model. We report ontologies, labeling results, model performances, and detailed analyses of the results. Our results show that the entities related to medications can be extracted with a relatively high accuracy of 0.90 F-score, followed by symptoms at 0.72 F-score, and conditions at 0.57 F-score. In our task, we not only identify where the symptoms are mentioned but also map them to canonical forms as they appear in the clinical notes. Of the different types of errors, in about 19-38% of the cases, we find that the model output was correct, and about 17-32% of the errors do not impact the clinical note. Taken together, the models developed in this work are more useful than the F-scores reflect, making it a promising approach for practical applications.

Laurent El Shafey, Hagen Soltau, Izhak Shafran

Speech applications dealing with conversations require not only recognizing the spoken words, but also determining who spoke when. The task of assigning words to speakers is typically addressed by merging the outputs of two separate systems, namely, an automatic speech recognition (ASR) system and a speaker diarization (SD) system. The two systems are trained independently with different objective functions. Often the SD systems operate directly on the acoustics and are not constrained to respect word boundaries and this deficiency is overcome in an ad hoc manner. Motivated by recent advances in sequence to sequence learning, we propose a novel approach to tackle the two tasks by a joint ASR and SD system using a recurrent neural network transducer. Our approach utilizes both linguistic and acoustic cues to infer speaker roles, as opposed to typical SD systems, which only use acoustic cues. We evaluated the performance of our approach on a large corpus of medical conversations between physicians and patients. Compared to a competitive conventional baseline, our approach improves word-level diarization error rate from 15.8% to 2.2%.

Hagen Soltau, Hank Liao, Hasim Sak

We present results that show it is possible to build a competitive, greatly simplified, large vocabulary continuous speech recognition system with whole words as acoustic units. We model the output vocabulary of about 100,000 words directly using deep bi-directional LSTM RNNs with CTC loss. The model is trained on 125,000 hours of semi-supervised acoustic training data, which enables us to alleviate the data sparsity problem for word models. We show that the CTC word models work very well as an end-to-end all-neural speech recognition model without the use of traditional context-dependent sub-word phone units that require a pronunciation lexicon, and without any language model removing the need to decode. We demonstrate that the CTC word models perform better than a strong, more complex, state-of-the-art baseline with sub-word units.

Tara N. Sainath, Brian Kingsbury, Abdel-rahman Mohamed et al.

Deep Convolutional Neural Networks (CNNs) are more powerful than Deep Neural Networks (DNN), as they are able to better reduce spectral variation in the input signal. This has also been confirmed experimentally, with CNNs showing improvements in word error rate (WER) between 4-12% relative compared to DNNs across a variety of LVCSR tasks. In this paper, we describe different methods to further improve CNN performance. First, we conduct a deep analysis comparing limited weight sharing and full weight sharing with state-of-the-art features. Second, we apply various pooling strategies that have shown improvements in computer vision to an LVCSR speech task. Third, we introduce a method to effectively incorporate speaker adaptation, namely fMLLR, into log-mel features. Fourth, we introduce an effective strategy to use dropout during Hessian-free sequence training. We find that with these improvements, particularly with fMLLR and dropout, we are able to achieve an additional 2-3% relative improvement in WER on a 50-hour Broadcast News task over our previous best CNN baseline. On a larger 400-hour BN task, we find an additional 4-5% relative improvement over our previous best CNN baseline.