We recently developed a joint speech and language model (SLM [1]) 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 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) models, we propose a retrieval augmented SLM (ReSLM) that overcomes this weakness. We first train a retriever to retrieve text entities given audio inputs. The retrieved entities are then added as text inputs to the underlying LLM to bias model predictions. We evaluated ReSLM on speech MultiWoz task (DSTC-11 Challenge), and found that the 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 speech tasks requiring custom contextual information or domain-specific entities.

Dian Yu, Mingqiu Wang, Yuan Cao, Izhak Shafran, Laurent Shafey, Hagen Soltau. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

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).

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.

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.

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%.

In this paper, we introduce Spear, an open source and extensible toolbox for state-of-the-art speaker recognition. This toolbox is built on top of Bob, a free signal processing and machine learning library. Spear implements a set of complete speaker recognition toolchains, including all the processing stages from the front-end feature extractor to the final steps of decision and evaluation. Several state-of-the-art modeling techniques are included, such as Gaussian mixture models, inter-session variability, joint factor analysis and total variability (i-vectors). Furthermore, the toolchains can be easily evaluated on well-known databases such as NIST SRE and MOBIO. As a proof of concept, an experimental comparison of different modeling techniques is conducted on the MOBIO database.

The problem of gender recognition using visual and acoustic cues has recently received significant attention. This paper explores the use of Total Variability (i-vectors) and Inter-Session Variability (ISV) modeling techniques for both unimodal and bimodal gender recognition, and compares them to several state-of-the-art algorithms. The experimental evaluation is conducted on the FERET and LFW databases for face-based gender recognition, on the NIST-SRE database for audio-based gender recognition, and on the MOBIO database for audio-visual gender recognition. Results on LFW show that the i-vectors technique outperforms state-of-the-art algorithms, which are based on Support Vector Machines (SVM) applied either on raw pixels, on Local Binary Patterns (LBP) or on Gabor filters, with an accuracy rate of about 95%. Results on NIST-SRE show that the i-vectors system is also superior to state-of-the-art GMM-based gender recognition systems, with a relative gain of about 11%. Finally, results on MOBIO show that i-vectors and ISV also take advantage of combining visual and acoustic cues using logistic regression. The resulting bimodal systems achieve accuracy rates of about 98%.

I4U is a joint entry of nine research Institutes and Universities across 4 continents to NIST SRE 2012. It started with a brief discussion during the Odyssey 2012 workshop in Singapore. An online discussion group was soon set up, providing a discussion platform for different issues surrounding NIST SRE’12. Noisy test segments, uneven multi-session training, variable enrollment duration, and the issue of open-set identification were actively discussed leading to various solutions integrated to the I4U submission. The joint submission and several of its 17 sub-systems were among top-performing systems. We summarize the lessons learnt from this large-scale effort.

In this paper, we present a scalable and exact solution for probabilistic linear discriminant analysis (PLDA). PLDA is a probabilistic model that has been shown to provide state-of-the-art performance for both face and speaker recognition. However, it has one major drawback: At training time estimating the latent variables requires the inversion and storage of a matrix whose size grows quadratically with the number of samples for the identity (class). To date, two approaches have been taken to deal with this problem, to 1) use an exact solution that calculates this large matrix and is obviously not scalable with the number of samples or 2) derive a variational approximation to the problem. We present a scalable derivation which is theoretically equivalent to the previous nonscalable solution and thus obviates the need for a variational approximation. Experimentally, we demonstrate the efficacy of our approach in two ways. First, on labeled faces in the wild, we illustrate the equivalence of our scalable implementation with previously published work. Second, on the large Multi-PIE database, we illustrate the gain in performance when using more training samples per identity (class), which is made possible by the proposed scalable formulation of PLDA.

Bob is a free signal processing and machine learning toolbox originally developed by the Biometrics group at Idiap Research Institute, Switzerland. The toolbox is designed to meet the needs of researchers by reducing development time and efficiently processing data. Firstly, Bob provides a researcher-friendly Python environment for rapid development. Secondly, efficient processing of large amounts of multimedia data is provided by fast C++ implementations of identified bottlenecks. The Python environment is integrated seamlessly with the C++ library, which ensures the library is easy to use and extensible. Thirdly, Bob supports reproducible research through its integrated experimental protocols for several databases. Finally, a strong emphasis is placed on code clarity, documentation, and thorough unit testing. Bob is thus an attractive resource for researchers due to this unique combination of ease of use, efficiency, extensibility and transparency. Bob is an open-source library and an ongoing community effort.