|
|||||||||||||||||||||||||||||||||
Feature Articles: Front-line of Speech, Language, and Hearing Research for Heartfelt Communications Advances in Multi-speaker Conversational Speech Recognition and UnderstandingAbstractOpportunities have been increasing in recent years for ordinary people to use speech recognition technology. For example, we can easily operate smartphones using voice commands. However, attempts to construct a device that can recognize human conversation have produced unsatisfactory results in terms of accuracy and usability because current technology is not designed for this purpose. At NTT Communication Science Laboratories, our goal is to create a new technology for multi-speaker conversational speech recognition and understanding. In this article, we review the technology we have developed and present our meeting analysis system that can accurately recognize who spoke when, what, to whom, and how in meeting situations. Keywords: multi-speaker, speech recognition, diarization 1. IntroductionA meeting is a basic human activity in which a group of people share information, present opinions, and make decisions. In formal meetings, it is standard for one person to take minutes. However, it often happens that certain important details are forgotten and therefore not recorded in the minutes. Moreover, meetings are not always easy to control, and this sometimes makes it difficult to achieve the objectives of the meeting. The participants may also be ill-informed, which can lead to misunderstandings or disagreements. Consequently, technology that is capable of automatically recognizing and understanding speech used in meetings has been attracting increasing attention [1], [2] as a way to overcome such problems. Today, speech recognition technology is widely used in many applications such as the operation of smartphones using voice commands. If we speak clearly into such a device, the spoken words can be recognized correctly and the command executed as intended. However, when we try to construct a device that can be applied to recognize conversations in meetings, as many as half of the words are not recognized correctly. This is because speech signals are often degraded by background noise and the voices of other participants, and conversational speech itself involves a wide variety of acoustic and linguistic patterns compared with speech directed at a device. As a result, the speech recognition accuracy deteriorates significantly. At NTT Communication Science Laboratories, we are working hard to develop meeting speech recognition technology that can solve these problems. However, another problem is that even if a speech recognizer achieves 100 percent accuracy for a meeting, no information about the meeting will be provided except for the spoken word sequence in a text format. This means that we can understand what words were spoken in the meeting but not who spoke when, to whom, and in what manner, which are all important pieces of information if we are to understand any meeting. In our research group, we are also studying meeting analysis technology that will enable us to understand a meeting in its entirety [2]. Our aim is to create a system that simultaneously obtains verbal information by speech recognition and nonverbal information by audio-visual scene analysis. We have already developed a prototype system for meeting analysis, which we designed to evaluate and demonstrate our proposed techniques. The first version of the system visualized a meeting based on non-verbal information, where the system recognized who spoke when and to whom and estimated the visual focus of attention using a microphone array and an omnidirectional camera [1]. We then extended the system to recognize both verbal and nonverbal information by incorporating our meeting speech recognition technology [2]. We have already shown that the system can both create draft meeting minutes and assist meeting participants with functions for looking back at past utterances and accessing information related to the words spoken during the meeting. In this article, we review the meeting speech recognition and understanding technology we have developed. In section 2, we describe our attempts to improve meeting speech recognition. In section 3, we present our meeting analysis system that accurately recognizes who spoke when, what, to whom, and how. We conclude the article and touch on future work in section 4. 2. Recognition of meeting speech2.1 Problems with meeting-speech recognitionWe consider an ordinary meeting room as shown in Fig. 1, where four meeting participants freely discuss various topics, and all the utterances are recorded with a microphone placed at the center of the table. However, speech recognition is not easy in this situation, and the recognition result will include many errors. There are two reasons for this problem, as described below.
(1) Conversation oblivious to microphones In face-to-face meetings, having participants wear a microphone or placing a microphone directly in front of each participant is not the preferred approach because it severely restricts the movement of the participants. Therefore, microphones should be located further away from each participant. However, this results in interference from acoustic noise and reverberation. Moreover, in settings where participants engage in informal and relaxed conversation, the utterances of two or more speakers often overlap. These factors significantly degrade speech recognition accuracy. (2) Acoustic and linguistic variety of spontaneous speech In private or informal meetings, people rarely speak formally and clearly. From acoustic and linguistic points of view, the utterances are fully spontaneous and therefore tend to include ambiguous pronunciations, abbreviations, and dialectal and emotional expressions. Consequently, a wide variety of speech patterns exist even for words that have the same dictionary pronunciation. These patterns change greatly depending on the speaker, the speaking style, and the topic. This aspect of spontaneous speech is also a crucial problem that degrades recognition accuracy. 2.2 Solutions to problemsTo solve these problems, we first worked on speech enhancement to improve the quality of speech signals in a meeting and proposed effective techniques based on a microphone array [2]. An overview of the speech enhancement method used in our meeting recognition system is shown in Fig. 2. The speech enhancement process consists of three phases: dereverberation, speaker diarization/source separation, and noise suppression.
(1) The dereverberation phase transforms the eight-channel microphone signals y1(τ)…y8(τ) into the time-frequency domain, removes reverberation components from the complex spectral sequences y1(f, t)…y8(f, t) of the microphone signals using multi-channel linear prediction, and outputs eight-channel dereverberated spectral sequences x1(f, t)…x8(f, t). (2) The speaker diarization/source separation phase detects active speakers based on direction of arrival (DOA) information, which is estimated by applying the generalized cross-correlation method with phase transform (GCC-PHAT) [3] to the dereverberated speech spectral sequences x1(f, t)…x8(f, t). In our system, the diarization result is obtained by clustering the DOAs at each frame, and the utterance period for each speaker n is output as the binary speaker diarization results d1(t)…d4(t), where dn(t) = 1 (or 0) indicates that speaker n is speaking (or silent) at frame t. Subsequently, source separation is performed to separate overlapping speech into speaker-dependent channels, where a null-beamforming approach is employed because it does not produce nonlinear artifacts that have detrimental effects on speech recognition. The beamformer coefficients for each speaker are estimated by leveraging the diarization result dn(t) [4]. (3) The noise suppression phase suppresses the noise components contained in each separated spectrum un(f, t), where we use Dynamic Integration of Voice Identification and DE-noising (DIVIDE) [5], [6]. DIVIDE reduces only the noise component of the original signals by using the online estimation of the speech and noise components, and then it outputs time-domain enhanced speech signals ŝn(τ). After the speech enhancement, speech recognition is performed by using the enhanced speech signals ŝn(τ), in which the diarization results dn(t) are also used to validate whether or not each recognized word is actually spoken by the participant associated with the separated channel. We recently improved the speech enhancement further by using DOLPHIN [7], which extracts the target speech more clearly based on the acoustic patterns of speech in the time and frequency domains. Although this method does not currently work with online processing, we confirmed that it yielded a large gain in recognition accuracy of meeting speech. 2.3 Automatic speech recognition moduleNext we present a brief overview of the automatic speech recognition (ASR) module that we designed for transcribing meeting speech. The module is based on SOLON [8], a speech recognizer that employs weighted finite-state transducers (WFSTs). SOLON employs an acoustic model consisting of a set of hidden Markov models (HMMs), a pronunciation lexicon, and language models represented as WFSTs that can be combined on the fly (i.e., as quickly as necessary) during decoding. The decoder efficiently finds the best scoring hypothesis in a search space organized with the given WFSTs. The input signal to the ASR module is spontaneous speech uttered by meeting participants, recorded with distant microphones, and enhanced by the audio processing techniques shown in Fig. 2. In general, it is effective to use a large amount of meeting speech data and their transcriptions to train acoustic and language models. However, there are no available Japanese data recorded under similar conditions, and it is very costly to collect new meeting data. Therefore, we prepared only a small amount of matched-condition data and used them to adapt the acoustic and language models. The acoustic model is a set of state-shared triphone HMMs, where each triphone (a sequence of three phonemes) is modeled as a left-to-right HMM with three states, and each shared state has a Gaussian mixture output distribution. First, initial HMMs are trained with a large corpus of clean speech data recorded via a close-talking microphone. The parameters of the initial HMMs are then estimated by discriminative training based on a differenced maximum mutual information (dMMI) criterion [9] to reduce recognition errors. Next, the initial HMMs are adapted with a small amount of real meeting data, which were recorded and enhanced with our meeting recognition system in advance. That is, the data were recorded with an 8-channel microphone array, and then dereverberated, separated, and subjected to noise suppression using the techniques in Fig. 2. The adaptation was performed by using maximum likelihood linear regression with automatically obtained multiple regression matrices [10]. We employ two types of language models. One is a standard back-off n-gram model. As with acoustic modeling, it is difficult to obtain a meeting transcript that is large enough to estimate the n-gram model. We use several types of data sets including a large written-text corpus and a small meeting transcript, and combine them with different weights based on an Expectation-Maximization algorithm. The other is a discriminative language model (DLM) trained with the R2D2 criterion [11]. This criterion is effective for training a language model that directly reduces recognition errors in a baseline speech recognizer. The decoder is based on efficient WFST-based one-pass decoding [8] in which fast on-the-fly composition can be used for combining WFSTs such as HCLG1 and G3/1 during decoding, where HCLG1 represents a WFST that transduces an HMM state sequence into a unigram-weighted word sequence, and G3/1 represents a WFST that weights a word sequence with the trigram probabilities divided by the unigram probabilities. This division is necessary to cancel out the unigram probabilities already contained in HCLG1. Since the algorithm can handle any number of WFSTs for composition on the fly, we combine four WFSTs, including two WFSTs that represent a DLM in the one-pass decoding. The first DLM WFST is based on word features and the second is based on part-of-speech (POS) features. Since a DLM can be represented as a set of word or POS n-grams with certain weights, it can be transformed into a WFST in the same way as a standard back-off n-gram model. The two DLM WFSTs for the word and POS features can be combined linearly by the composition operation during decoding. Thus, we can achieve one-pass real-time speech recognition using these DLM WFSTs unlike the conventional approaches that involve a rescoring step after the first-pass decoding. As part of the recent advances made in speech recognition technology in our research group, we have proposed an all-in-one speech recognition model [12], which is a different approach from those in which acoustic and language models are separately trained. The all-in-one model is represented as a WFST (or a set of WFSTs) including all the acoustic, pronunciation, and language models, and it is effectively trained using a discriminative criterion to reduce the number of recognition errors. As a result, the model can capture not only the general characteristics of acoustic and linguistic patterns but also a wide variety of interdependencies between acoustic and linguistic patterns in conversational speech. With this model, we have improved the speech recognition accuracy for meeting speech, and have increased the accuracy substantially by integrating the model with deep learning techniques [13]. 2.4 Diarization-based word filteringFinally, we describe diarization-based word filtering (DWF), which is an important technique used with our meeting speech recognizer. As shown in Fig. 2, we employed speech recognition for each speaker’s channel given by source separation. This approach is much more robust in the case of overlapping speech than that using only a single channel. However, even if we employ source separation, it is difficult to completely remove other speakers’ voices from the target speaker’s channel. Such remaining non-target speech signals often induce insertion errors in speech recognition. To solve this problem, we utilize frame-based speaker diarization results obtained using the method shown in Fig. 2 to reduce the number of insertion errors. Since the diarization result at each frame tends to be discontinuous on the time axis, we use the average value of the diarization results for each recognized word, which can be considered to represent the relevance of the word to the target speaker. With this measure, words with a low relevance can be effectively deleted from the recognition results. The relevance of word wn recognized for speaker n is computed as: where dn(t) is the frame-based diarization result at frame t, and b(wn) and e(wn) respectively indicate the beginning and ending frames of word wn. If s(wn) is less than a predefined threshold, wn is deleted from the speech recognition result. This method is effective for low-latency processing. Conventional methods detect a speech segment by speaker diarization, and then apply speech recognition for the segment. This requires a long time delay because speech recognition cannot start until the diarization step is finished. The DWF approach only requires a one-frame (32 msec) delay to obtain dn(t), and it can drive speech recognition in parallel. 3. Real-time meeting analysis systemOur meeting analyzer basically recognizes who is speaking what by using speech recognition and speaker diarization, and it detects the activity of each participant (e.g., speaking, laughing, looking at someone else) and the circumstances of the meeting (e.g., topic, activeness, casualness, and intelligibility (defined more specifically below)) by integrating results obtained from several processing modules. The detected results provide speaking to whom and how information. The results of analysis are continuously displayed on a browser running on an Android™* tablet, as shown in Fig. 3. The panel on the left side of the browser displays live streaming video of a 360-degree view provided by a camera together with information about who is speaking to whom represented by orange circles and light blue arrows. The right side shows the real-time transcript for each participant, as well as his/her picture. The face icon beside the transcript indicates the participant’s state (speaking, laughing, or silent), and the next two bars show the number of words spoken by the participant and how much he/she has been watched by others, i.e., the visual focus of attention for each speaker, based on the direction of each participant’s face. The lower left panel shows the current circumstances of the meeting in terms of topic words and the degrees of activeness, casualness, and intelligibility. Activeness is calculated as the number of words spoken multiplied by the entropy based on the relative frequencies of spoken words for all the speakers in a fixed time window. Casualness is estimated based on the frequency of laughter. Intelligibility is calculated based on the frequency of participants’ nods. These graphical representations help us to understand the current circumstances of the meeting both visually and objectively. The system architecture is depicted in Fig. 4. In the speech enhancement block, the m-th microphone signal in the STFT domain ym(f, t) (m = 1, …, 8) is dereverberated, separated, and denoised, and finally the enhanced signals ŝn(τ) are sent to the speech recognizer and the acoustic event detector. Here, f and t are frequency and time-frame indices, respectively. Speech recognition is used for each separated signal to transcribe utterances. The speech recognizer SOLON is also used to detect acoustic events including silence, speech, and laughter. Sentence boundary detection and topic tracking are applied to the word sequences from SOLON. In topic tracking, we use the Topic Tracking Language Model [14], which is an online extension of latent Dirichlet allocation that can adaptively track changes in topics by considering the information history of the meeting. In our system, we use conditional random fields for sentence boundary detection [15], where the transition features consist of bigrams of the labels that identify the presence or not of a sentence head. The other features consist of words and their POS tags in the scope of a 3-word context and the pause duration at each boundary candidate. The camera captures a 360-degree view from the center of the table. During visual processing, the faces of the participants are detected, and face images are sent to the browser at the beginning of the meeting. Then the face pose tracker continues to work during the meeting. This incorporates a Sparse Template Condensation Tracker [1], which realizes the real-time robust tracking of multiple faces by utilizing GPUs (graphics processing units). With this tracking approach, the position of each participant and his/her face direction can be obtained continuously. This visual information is used to determine who is speaking to whom and to detect the visual focus of attention. The position information is also used to associate each utterance with the corresponding participant by combining it with the DOA information of the speech signal. Accordingly, a transcript and a face icon can be displayed on the right panel for the participant who is speaking. The meeting analysis module calculates the activeness and casualness of the meeting based on speech recognition and acoustic event detection results. The intelligibility is obtained based on the frequency of participants’ nods detected by face-pose tracking. All the analysis results are sent to the browser together with streaming video via a real-time messaging protocol (RTMP) server. Since the RTMP server can receive multiple requests, the analysis results can be broadcast to multiple browsers. The current system runs on four computers: (1) an AMD Opteron 1389 2.9-GHz Quad Core for speaker diarization and source separation, (2) an Intel Xeon X5570 2.93 GHz 8-core dual processor for dereverberation and noise suppression, (3) the same Xeon model for speech recognition, acoustic event detection, and meeting analysis, and (4) an Intel Core 2 Extreme QX9650 3.0-GHz (+ NVIDIA GeForce9800GX2/2 GPU cores) for visual processing.
4. ConclusionIn this article, we reviewed the technology we have developed in our research group and presented our meeting analysis system that provides accurate recognition of who spoke when, what, to whom, and how in a meeting situation. If conversational speech recognition and understanding based on audio-visual scene analysis becomes possible, many useful applications could be realized. In the future, it may be possible not only to generate meeting minutes automatically, but also to easily find past meeting scenes when required. We might have a virtual secretary who could answer our questions and register our plans in the scheduler autonomously. To realize such a system, it is important to improve speech recognition accuracy, and also to detect what is occurring in the surroundings, how the speaker is feeling, and why the meeting led us to a certain conclusion, etc. To enable such a deep understanding of human conversation, we need to extend the meeting analysis technology so that it can recognize higher-level concepts. To this end, we are continuing to address problems beyond the framework of speech recognition technology. References
|