In this paper, we propose a maximum entropy based approach for automatic annotation of multimedia content. In our approach, we explicitly model the spatial-location of the low-level features by means of specially designed predicates. In addition, the interaction between the low-level features is modeled using joint observation predicates. We evaluate the performance of semantic concept classifiers built using this approach on the TRECVID2003 corpus. Experiments indicate that our model performance is on par with the best results reported to-date on this dataset; despite using only unimodal features and a single approach towards model-building. This compares favorably with the state-of-the-art systems which use multimodal features and classifier fusion to achieve similar results on this corpus.

In this paper we describe a novel approach for jointly modeling the text and the visual components of multimedia documents for the purpose of information retrieval(IR). We propose a novel framework where individual components are developed to model different relationships between documents and queries and then combined into a joint retrieval framework. In the state-of-the-art systems, a late combination between two independent systems, one analyzing just the text part of such documents, and the other analyzing the visual part without leveraging any knowledge acquired in the text processing, is the norm. Such systems rarely exceed the performance of any single modality (i.e. text or video) in information retrieval tasks. Our experiments indicate that allowing a rich interaction between the modalities results in significant improvement in performance over any single modality. We demonstrate these results using the TRECVID03 corpus, which comprises 120 hours of broadcast news videos. Our results demonstrate over 14 % improvement in IR performance over the best reported text-only baseline and ranks amongst the best results reported on this corpus.

This paper describes multimodal systems for ad-hoc search constructed by IBM for the TRECVID 2003 benchmark of search systems for broadcast video. These systems all use a late fusion of independently developed speech-based and visual content-based retrieval systems and outperform our individual retrieval systems on both manual and interactive search tasks. For the manual task, our best system used a query-dependent linear weighting between speech-based and image-based retrieval systems. This system has mean average precision (MAP) performance 20% above our best unimodal system for manual search. For the interactive task, where the user has full knowledge of the query topic and the performance of the individual search systems, our best system used an interlacing approach. The user determines the (subjectively) optimal weights A and B for the speech-based and image-based systems, where the multimodal result set is aggregated by combining the top A documents from system A followed by top B documents of system B and then repeating this process until the desired result set size is achieved. This multimodal interactive search has MAP 40% above our best unimodal interactive search system.

Visual information in a speaker's face is known to improve the robustness of automatic speech recognition (ASR). However, most studies in audio-visual ASR have focused on "visually clean" data to benefit ASR in noise. This paper is a follow up on a previous study that investigated audio-visual ASR in visually challenging environments. It focuses on visual speech front end processing, and it proposes an improved, appearance based face and feature detection algorithm that utilizes Gaussian mixture model classifiers. This method is shown to improve the accuracy of face and feature detection, and thus visual speech recognition, over our previously used baseline system. In turn, this translates to improved audio-visual ASR, resulting in a 10% relative reduction of the word-error-rate in noisy speech.

We present our approach to detecting monologues in video shots. A monologue shot is defined as a shot containing a talking person in the video channel with the corresponding speech in the audio channel. Whilst motivated by the TREC 2002 Video Retrieval Track (VT02), the underlying approach of synchrony between audio and video signals is also applicable for voice and face-based biometrics, assessing lip-synchronization quality in movie editing, and for speaker localization in video. Our approach is envisioned as a two part scheme. We first detect the occurrence of speech and face in a video shot. In shots containing both speech and a face, we distinguish monologue shots as those shots where the speech and facial movements are synchronized. To measure the synchrony between speech and facial movements we use a mutual-information based measure. Experiments with the VT02 corpus indicate that using synchrony, the average precision improves by more than 50% relative compared to using face and speech information alone. Our synchrony based monologue detector submission had the best average precision performance (in VT02) amongst 18 different submissions.

In this paper we present our approach to detect monologues in video shots. A monologue shot is defined as a shot containing a talking person in the video channel with the corresponding speech in the audio channel. Whilst motivated by the TREC 2002 video retrieval track (VT02), the underlying approach of synchrony between audio and video signals are also applicable for voice and face-based biometrics, assessing of lip-synchronization quality in movie editing, and for speaker localization in video. Our approach is envisioned as a two part scheme. We first detect occurrence of speech and face in a video shot. In shots containing both speech and a face, we distinguish monologue shots as those shots where the speech and facial movements are synchronized. To measure the synchrony between speech and facial movements we use a mutual-information based measure. Experiments with the VT02 corpus indicate that using synchrony, the average precision improves by more than 50% relative compared to using face and speech information alone. Our synchrony based monologue detector submission had the best average precision performance (in VT02) amongst 18 different submissions.

This paper addresses the influence of audio segmentation and segment clustering on automatic transcription accuracy for large spoken archives. The work forms part of the ongoing MALACH project, which is developing advanced techniques for supporting access to the world’s largest digital archive of video oral histories collected in many languages from over 52000 survivors and witnesses of the Holocaust. We present several audio-only and audio-visual segmentation schemes, including two novel schemes: the first is iterative and audio-only, the second uses audio-visual synchrony. Unlike most previous work, we evaluate these schemes in terms of their impact upon recognition accuracy. Results on English interviews show the automatic segmentation schemes give performance comparable to (exhorbitantly expensive and impractically lengthy) manual segmentation when using a single pass decoding strategy based on speaker-independent models. However, when using a multiple pass decoding strategy with adaptation, results are sensitive to both initial audio segmentation and the scheme for clustering segments prior to adaptation: the combination of our best automatic segmentation and clustering scheme has an error rate 8% worse (relative) to manual audio segmentation and clustering due to the occurrence of “speaker-impure ” segments. 1.

In this paper we describe a general information fusion algorithm that can be used to incorporate multimodal cues in building user-defined semantic concept models. We compare this technique with a Bayesian Network-based approach on a semantic concept detection task. Results indicate that this technique yields superior performance. We demonstrate this approach further by building classifiers of arbitrary concepts in a score space defined by a pre-deployed set of multimodal concepts. Results show annotation for user-defined concepts both in and outside the pre-deployed set is competitive with our best video-only models on the TREC Video 2002 corpus.

This paper describes progress towards a general framework for incorporating multimodal cues into a trainable system for automatically annotating user-defined semantic concepts in broadcast video. Models of arbitrary concepts are constructed by building classifiers in a score space defined by a pre-deployed set of multimodal models. Results show annotation for user-defined concepts both in and outside the pre-deployed set is competitive with our best video-only models on the TREC Video 2002 corpus. An interesting side result shows speech-only models give performance comparable to our best video-only models for detecting visual concepts such as "outdoors", "face" and "cityscape".

We describe methods for automatic labeling of high-level semantic concepts in documentary style videos. The emphasis of this paper is on audio processing and on fusing information from multiple modalities. The work described represents initial work towards a trainable system that acquires a collection of generic "intermediate" semantic concepts across modalities (such as audio, video, text) and combines information from these modalities for automatic labeling of a "high-level" concept. Initial results suggest that multi-modal fusion achieves a 12.5% relative improvement over the best unimodal model.

This paper provides a detailed description of the University of Washington automatic speech recognition (ASR) system for the 2001 DARPA SPeech In Noisy Environments (SPINE) task. Our system makes heavy use of the graphical modeling toolkit (GMTK), a general purpose graphical modeling-based ASR system that allows arbitrary parameter tying, flexible deterministic and stochastic dependencies between variables, and a generalized maximum likelihood parameter estimation algorithm. In our SPINE system, GMTK was used for acoustic model training whereas feature extraction, speaker adaptation, and first-pass decoding were performed by HTK. Our integrated GMTK/HTK system demonstrates the relative merits provided by each tool. Novel aspects of our SPINE system include the capturing of correlations among feature vectors via a globally-shared factored sparse inverse covariance matrix and generalized EM training. 1.

This paper considers schemes for determining which of a set of faces on screen, if any, is producing speech in a video soundtrack. Whilst motivated by the TREC 2002 (Video Retrieval Track) monologue detection task, the schemes are also applicable to voice and face-based biometrics systems, for assessing lip synchronization quality in movie editing and computer animation, and for speaker localization in video. Several approaches are discussed: two implementations of a generic mutual-information-based measure of the degree of synchrony between signals, which can be used with or without prior speech and face detection, and a stronger model-based scheme which follows speech and face detection with an assessment of face and lip movement plausibility. Schemes are compared on a corpus of 1016 test cases containing multiple faces and multiple speakers, a test set 200 times larger than the nearest comparable test set of which we are aware. The most successful and computationally cheapest scheme obtains an accuracy of 82% on the task of picking the consistent speaker from a set including three confusers. A final experiment demonstrates the potential utility of the scheme for speaker localization in video.

Hidden Markov Models (HMMs) have been successful for modelling the dynamics of carefully dictated speech, but their performance degrades severely when used to model conversational speech. This paper presents a preliminary feasibility study of an alternative class of models: loosely coupled HMMs. Since speech is produced by a system of loosely coupled articulators, stochastic models explicitly representing this parallelism may have advantages for automatic speech recognition (ASR), particularly when trying to model the phonological effects inherent in casual spontaneous speech. The paper evaluates one coupled model on a simple ASR task, using both exact and approximate estimation schemes. We conclude such models merit further investigation. 1. INTRODUCTION Hidden Markov Models (HMMs) have been successful for modelling the dynamics of carefully dictated speech. However, their performance degrades severely when they are used to model conversational speech, and it has been widely hypothe...

Conversational speech exhibits considerable pronunciation variability, which has been shown to have a detrimental effect on the accuracy of automatic speech recognition. There have been many attempts to model pronunciation variation, including the use of decision trees to generate alternate word pronunciations from phonemic baseforms. Use of pronunciation models during recognition is known to improve accuracy. This paper describes the incorporation of pronunciation models into acoustic model training in addition to recognition. Subtle difficulties in the straightforward use of alternatives to canonical pronunciations are first illustrated: it is shown that simply improving the accuracy of the phonetic transcription used for acoustic model training is of little benefit. Acoustic models trained on the most accurate phonetic transcriptions result in worse recognition than acoustic models trained on canonical baseforms. Analysis of this counterintuitive result leads to a new method of accommodating nonstandard pronunciations: rather than allowing a phoneme in the canonical pronunciation to be realized as one of a few distinct alternate phones, the hidden Markov model (HMM) states of the phoneme’s model are instead allowed to share Gaussian mixture components with the HMM states of the model(s) of the alternate realization(s). Qualitatively, this amounts to making a soft decision about which surface form is realized. Quantitatively, experiments show that this method is particularly well suited for acoustic model training for spontaneous speech: a 1.7 %(absolute) improvement in recognition accuracy on the Switchboard corpus is presented.

Conversational speech exhibits considerable pronunciation variability, which has been shown to have a detrimental effect on the accuracy of automatic speech recognition. There have been many attempts to model pronunciation variation, including the use of decision trees to generate alternate word pronunciations from phonemic baseforms. Use of pronunciation models during recognition is known to improve accuracy. This paper describes the incorporation of pronunciation models into acoustic model training in addition to recognition. Subtle difficulties in the straightforward use of alternatives to canonical pronunciations are first illustrated: it is shown that simply improving the accuracy of the phonetic transcription used for acoustic model training is of little benefit. Acoustic models trained on the most accurate phonetic transcriptions result in worse recognition than acoustic models trained on canonical baseforms. Analysis of this counterintuitive result leads to a new method of accommodating nonstandard pronunciations: rather than allowing a phoneme in the canonical pronunciation to be realized as one of a few distinct alternate phones, the hidden Markov model (HMM) states of the phoneme?s model are instead allowed to share Gaussian mixture components with the HMM states of the model(s) of the alternate realization(s). Qualitatively, this amounts to making a soft decision about which surface form is realized. Quantitatively, experiments show that this method is particularly well suited for acoustic model training for spontaneous speech: a 1.7 %(absolute) improvement in recognition accuracy on the Switchboard corpus is presented.

Accurately modelling pronunciation variability in conversational speech is an important component of an automatic speech recognition system. We describe some of the projects undertaken in this direction during and after WS97, the Fifth LVCSR Summer Workshop, held at Johns Hopkins University, Baltimore, in July-August, 1997. We first illustrate a use of hand-labelled phonetic transcriptions of a portion of the Switchboard corpus, in conjunction with statistical techniques, to learn alternatives to canonical pronunciations of words. We then describe the use of these alternate pronunciations in an automatic speech recognition system. We demonstrate that the improvement in recognition performance from pronunciation modelling persists as the system is enhanced with better acoustic and language models.

Phonetic decision trees have been widely used for obtaining robust context-dependent models in HMM-based systems. There are five key issues to consider when constructing phonetic decision trees: the alignment of data with the chosen phone classes; the quality of the modeling of the underlying data; the choice of partitioning method at each node; the goodness-of-split criterion and the method for determining appropriate tree sizes. A popular existing method usesefficient but crude approximatemethods for each of these. This paper introduces and evaluates more detailed alternatives to the standard approximations. 1. Introduction A key problem in building continuous-density Hidden Markov Model (HMM)-based context-dependent acoustic models is maintaining a balancebetween the desired model complexity and the number of parameters which can be robustly estimated from the available training data. One solution which has proved successful (eg.[8], [1]) is based upon the use of phonetic decision...