Search

This paper describes a three-year project at the National Research Council of Canada aimed at developing software to assist Indigenous communities in their efforts to preserve their languages and extend their use. The project aimed to work within the empowerment paradigm, where the linguistic goals of communities have at least equal weight with those of the researchers, and where collaboration with communities is central. Because many of the technological directions we took were in response to community needs, the project ended up as a collection of diverse subprojects, including the creation of a sophisticated framework for building verb conjugators for highly inflectional polysynthetic languages (a verb conjugator for Kanyen’kéha, in the Iroquoian language family, was built in the framework), release of what is probably the largest available corpus of sentences in a polysynthetic language (Inuktut) aligned with English sentences and experiments with machine translation (MT) systems trained on this corpus, free online services based on automatic speech recognition (ASR) for easing the transcription bottleneck for recordings of speech in Indigenous languages (and other languages), limited-domain text-to-speech synthesis for some Indigenous languages, and several other subprojects., Ce rapport technique décrit un projet de trois ans au Conseil national de recherches du Canada pour le développement de logiciels visant à soutenir les communautés autochtones dans leurs efforts de préservation et de revitalisation de leurs langues. Ce projet s’efforce de respecter le principe de l’autonomisation : l’importance accordée aux objectifs linguistiques des communautés est supérieure ou égale à celle accordée à ceux des chercheurs et la collaboration avec les communautés est centrale. Puisque beaucoup des orientations technologiques ont été choisies en réponse aux besoins des communautés, le projet a donné lieu à divers sous-projets, notamment : la création d’une plateforme pour créer des conjugueurs pour les langues polysynthétiques hautement flexionnelles (un conjugueur pour la langue iroquoienne kanyen'kéha a été développé à l’aide de cette plateforme); la publication d’un corpus bilingue inuktut–anglais, qui est probablement le plus grand corpus de phrases disponible pour une langue polysynthétique; un projet d’expérimentation de la traduction automatique entrainée sur ce corpus; des services en ligne gratuits basés sur la reconnaissance de la parole automatique pour soulager le goulot d’étranglement de la transcription des enregistrements en langues autochtones; un projet sur l’utilisation de la synthèse vocale à domaine restreint pour certaines langues autochtones; et plusieurs autres sous-projets.

In this paper, we present our audio fingerprinting system that detects a transformed copy of an audio from a large collection of audios in a database. The audio fingerprints in this system encode the positions of salient regions of binary images derived from a spectrogram matrix. The similarity between two fingerprints is defined as the intersection of their elements (i.e. positions of the salient regions). The search algorithm labels each reference fingerprint in the database with the closest query frame and then counts the number of matching frames when the query is overlaid over the reference. The best match is based on this count. The salient regions fingerprints together with this nearest-neighbor search give excellent copy detection results. However, for a large database, this search is time consuming. To reduce the search time, we accelerate this similarity search by using a graphics processing unit (GPU). To speed this search even further, we use a two-step search based on a clustering technique and a lookup table that reduces the number of comparisons between the query and the reference fingerprints. We also explore the tradeoff between the speed of search and the copy detection performance. The resulting system achieves excellent results on TRECVID 2009 and 2010 datasets and outperforms several state-of-the-art audio copy detection systems in detection performance, localization accuracy and run time. For a fast detection scenario with detection speed comparable to the Ellis' Shazam-based system, our system achieved the same min NDCR as the NN-based system, and significantly better detection accuracy than Ellis' Shazam-based system.

This paper presents a novel audio fingerprinting method that is highly robust to a variety of audio distortions. It is based on an unconventional audio fingerprint generation scheme. The robustness is achieved by generating different versions of the spectrogram matrix of the audio signal by using a threshold based on the average of the spectral values to prune this matrix. We transform each version of this pruned spectrogram matrix into a 2-D binary image. Multiple versions of these 2-D images suppress noise to a varying degree. This varying degree of noise suppression improves likelihood of one of the images matching a reference image. To speed up matching, we convert each image into an n-dimensional vector, and perform a nearest neighbor search based on this n-dimensional vector. We give results with two different feature parameters and their combination. We test this method on TRECVID 2010 content-based copy detection evaluation dataset, and we validate the performance on TRECVID 2009 dataset also. Experimental results show the effectiveness of these features even when the audio is distorted. We compare the proposed method to two state-of-the-art audio copy detection systems, namely NN-based and Shazam systems. Our method by far outperforms Shazam system for all audio transformations (or distortions) in terms of detection performance, number of missed queries and localization accuracy. Compared to NN-based system, our approach reduces minimal Normalized Detection Cost Rate (min NDCR) by 23 % and improves localization accuracy by 24 %.

This paper describes a video fingerprinting system that is highly robust to audio and video transformations. The proposed system adapts a robust audio fingerprint extraction approach to video fingerprinting. The audio fingerprinting system converts the spectrogram into binary images, and then encodes the positions of salient regions selected from each binary image. Visual features are extracted in a similar way from the video images. We propose two visual fingerprint generation methods where fingerprints encode the positions of salient regions of greyscale video images. Salient regions of the first method are selected based on the intensity values of the image, while the second method identifies the regions that represent the highest variations between two successive images. The similarity between two fingerprints is defined as the intersection between their elements. The search algorithm is speeded up by an efficient implementation on a Graphics Processing Unit (GPU). We evaluate the performance of the proposed video system on TRECVID 2009 and 2010 datasets, and we show that this system achieves promising results and outperforms other state-of-the-art video copy detection methods for queries that do not includes geometric transformations. In addition, we show the effectiveness of this system for a challenging audio+video copy detection task.

We propose to improve the performance of i-vector based speaker verification by processing the i-vectors with a deep neural network before they are fed to a cosine distance or probabilistic linear discriminant analysis (PLDA) classifier. To this end we build on an existing model that we refer to as Non-linear Within Class Normalization (NWCN) and introduce a novel Speaker Classifier Network (SCN). Both models deliver impressive speaker verification performance, showing a 56% and 68% relative improvement over standard i-vectors when combined with a cosine distance backend. The NWCN model also reduces the equal error rate for PLDA from 1.78% to 1.63%. We also test these models under the constraints of domain mismatch, i.e. when no in-domain training data is available. Under these conditions, SCN features in combination with cosine distance performs better than the PLDA baseline, achieving an equal error rate of 2.92% as compared to 3.37%.

We report results on speaker diarization of telephone conversations. This speaker diarization process is similar to the multistage segmentation and clustering system used in broadcast news. It consists of an initial acoustic change point detection algorithm, iterative Viterbi re-segmentation, gender labeling, agglomerative clustering using a Bayesian information criterion (BIC), followed by agglomerative clustering using state-of-the-art speaker identification (SID) methods and Viterbi re-segmentation using Gaussian mixture models (GMMs). We repeat these multistage segmentation and clustering steps twice: once with mel-frequency cepstral coefficients (MFCCs) as feature parameters for the GMMs used in gender labeling, SID, and Viterbi re-segmentation steps and another time with Gaussianized MFCCs as feature parameters for the GMMs used in these three steps. The resulting clusters from the parallel runs are combined in a novel way that leads to a significant reduction in the diarization error rate (DER). On a development set containing 30 telephone conversations, this combination step reduced the DER by 20%. On another test set containing 30 telephone conversations, this step reduced the DER by 13%. The best error rate we have achieved is 6.7% on the development set and 9.0% on the test set.

The REVERB challenge provides a common framework for the evaluation of feature extraction techniques in the presence of both reverberation and additive background noise. State-of-the-art speech recognition systems perform well in controlled environments, but their performance degrades in realistic acoustical conditions, especially in real as well as simulated reverberant environments. In this contribution, we utilize multiple feature extractors including the conventional mel-filterbank, multi-taper spectrum estimation-based mel-filterbank, robust mel and compressive gammachirp filterbank, iterative deconvolution-based dereverberated mel-filterbank, and maximum likelihood inverse filtering-based dereverberated mel-frequency cepstral coefficient features for speech recognition with multi-condition training data. In order to improve speech recognition performance, we combine their results using ROVER (Recognizer Output Voting Error Reduction). For two- and eight-channel tasks, to get benefited from the multi-channel data, we also use ROVER, instead of the multi-microphone signal processing method, to reduce word error rate by selecting the best scoring word at each channel. As in a previous work, we also apply i-vector-based speaker adaptation which was found effective. In speech recognition task, speaker adaptation tries to reduce mismatch between the training and test speakers. Speech recognition experiments are conducted on the REVERB challenge 2014 corpora using the Kaldi recognizer. In our experiments, we use both utterance-based batch processing and full batch processing. In the single-channel task, full batch processing reduced word error rate (WER) from 10.0 to 9.3 % on SimData as compared to utterance-based batch processing. Using full batch processing, we obtained an average WER of 9.0 and 23.4 % on the SimData and RealData, respectively, for the two-channel task, whereas for the eight-channel task on the SimData and RealData, the average WERs found were 8.9 and 21.7 %, respectively.

This paper presents the latest improvements on our Spectro system that detects transformed duplicate audio content. We propose a new binary image feature derived from a spectrogram matrix by using a threshold based on the average of the spectral values. We quantize this binary image by applying a tile of fixed size and computing the sum of each small square in the tile. Fingerprints of each binary image encode the positions of the selected tiles. Evaluation on TRECVID 2010 CBCD data shows that this new feature improves significantly the Spectro system for transformations that add irrelevant speech to the audio. Compared to a state-of-the-art audio fingerprinting system, the proposed method reduces the minimal Normalized Detection Cost Rate (min NDCR) by 33%, improves localization accuracy by 28% and results in 40% fewer missed queries.

We investigate the use of deep neural nets (DNN) to provide initial speaker change points in a speaker diarization system. The DNN trains states that correspond to the location of the speaker change point (SCP) in the speech segment input to the DNN. We model these different speaker change point locations in the DNN input by 10 to 20 states. The confidence in the SCP is measured by the number of frame synchronous states that correspond to the hypothesized speaker change point. We only keep the speaker change points with the highest confidence. We show that this DNN-based change point detector reduces the number of missed change points for both an English test set and a French dev set. We also show that the DNN-based change points reduce the diarization error rate for both an English and a French diarization system. These results show the feasibility of DNNs to provide initial speaker change points.

The Multi-Genre Broadcast Challenge at ASRU 2015 is a controlled evaluation of speech recognition, speaker diarization, and lightly supervised alignment using BBC TV recordings. CRIM and LIUM teams participated in the speech recognition part of the challenge with a joint submission. This paper presents the CRIM and LIUM's contributions. Each team made different choices to develop its ASR system. By the way, it was expected to compare and to evaluate different approaches to diarization and acoustic modeling, and to get complementary ASR systems for effective merging. CRIM's main contributions are the use of a training scenario similar to multi-lingual training to estimate the deep neural net (DNN) acoustic models with most of the data, the use of a pruned trigram model for search, in addition to the use of a genre-dependent quadgram language model for rescoring the lattice from the search. For LIUM, the focus was on fast decoding with high accuracy. The final word error rates (WER) after merging show that it is possible to get reasonable WER with automatically aligned files. The final global WER of 25.1% corresponds to a WER reduction of about 20% absolute in comparison to the ASR baseline system provided by the organizers.

This paper presents a novel audio fingerprinting method that is highly robust to a variety of audio distortions. It is based on unconventional audio fingerprints generation scheme. The robustness is achieved by generating different versions of the spectrogram matrix of the audio signal by using a threshold based on the average of the spectral values to prune this matrix. We transform each version of this pruned spectrogram matrix into a 2-D binary image. Multiple 2-D images suppress noise to a varying degree. This varying degree of noise suppression improves likelihood of one of the images matching a reference image. To speed up matching, we convert each image into an n-dimensional vector, and perform a nearest neighbor search based on this n-dimensional vector. We test this method on TRECVID 2010 content-based copy detection evaluation dataset. Experimental results show the effectiveness of such fingerprints even when the audio is distorted. We compare the proposed method to a state-of-the-art audio copy detection system. Results of this comparison show that our method achieves an improvement of 22% in localization accuracy, and lowers minimal normalized detection cost rate (min NDCR) by half for audio transformations T1 and T2.

In this paper we look at real-time computing issues in large vocabulary speech recognition. We use the French broadcast audio transcription task from ETAPE 2011 for this evaluation. We compare word error rate (WER) versus overall computing time for hidden Markov models with Gaussian mixtures (GMM-HMM) and deep neural networks (DNN-HMM). We show that for a similar computing during recognition, the DNNHMM combination is superior to the GMM-HMM. For a realtime computing scenario, the error rate for the ETAPE dev set is 23.5% for DNN-HMM versus 27.9% for the GMM-HMM: a significant difference in accuracy for comparable computing. Rescoring lattices (generated by DNN-HMM acoustic model) with a quadgram language model (LM), and then with a neural net LM reduces the WER to 22.0% while still providing realtime computing.

In this paper, we introduce the concept of the effective sample size to speaker diarization and recognition. We show why the use of the nominal sample size is inadequate to feature streams that exhibit inter-frame correlations and how it adversely affects inference. We then discuss the effective sample size, that is the sample size of a set of independent observations that carry the equivalent amount of statistical information about the model parameters and how the scaling factor can be estimated. Our experiments on speaker diarization show that once the effective sample size is adopted, state-of-the-art results can be attained even with single Gaussians and Hierarchical Clustering, and even when the scaling factor is set to be common for all utterances. On speaker recognition, encouraging results are reported on NIST-2010 using iVectors and PLDA.

We report results on video copy detection using nearest-neighbor (NN) mapping that has been used successfully in audio copy detection. For copy detection search, we use a sliding window to move the query video over the test video, and count the number of frames of query that match the frames in the test segment. The feature in the test frame that we match is the frame number of the query that is closest to that test frame. This leads to good matching scores even when the query video is distorted and contains occlusions. We test the NN mapping algorithm and the video features that map test frame to the closest query frame on TRECVID 2009 and 2010 content-based copy detection (CBCD) evaluation data. For both these tasks, the NN mapping for video copy detection gives minimal normalized detection cost rate (min NDCR) comparable to that achieved with audio copy detection for the same task. For the TRECVID 2011 CBCD evaluation data we got the lowest min NDCR for 26 out of 56 transforms for actual no false alarm case.

We present an application of video indexing/summarization to produce Videodescription (VD) for the blinds. Audio and computer vision technologies can automatically detect and recognize many elements that are pertinent to VD which can speed-up the VD production process. We have developed and integrated many of them into a first computer-assisted VD production software. The paper presents the main outcomes of this R&D activity started 5 years ago in our laboratory. Up to now, usability performance on various video and TV series types have shown a reduction of up to 50% in the VD time production process.

This paper investigates spoken term detection (STD) from audio recordings of course lectures obtained from an existing media repository. STD is performed from word lattices generated offline using an automatic speech recognition (ASR) system configured from a meetings domain. An efficient STD approach is presented where lattice paths which are likely to contain search terms are identified and an efficient phone based distance is used to detect the occurrence of search terms in phonetic expansions of promising lattice paths. STD and ASR results are reported for both in-vocabulary (IV) and out-of-vocabulary (OOV) search terms in this lecture speech domain.

The authors demonstrate the effectiveness of phonemic hidden Markov models with Gaussian mixture output densities (mixture HMMs) for speaker-dependent large-vocabulary word recognition. Speech recognition experiments show that for almost any reasonable amount of training data, recognizers using mixture HMMs consistently outperform those employing unimodal Gaussian HMMs. With a sufficiently large training set (e.g. more than 2500 words), use of HMMs with 25-component mixture distributions typically reduces recognition errors by about 40%. It is also found that the mixture HMMs outperform a set of unimodal generalized triphone models having the same number of parameters. Previous attempts to employ mixture HMMs for speech recognition proved discouraging because of the high complexity and computational cost in implementing the Baum-Welch training algorithm. It is shown how mixture HMMs can be implemented very simply in unimodal transition-based frameworks by allowing multiple transitions from one state to another. >

We report results on speaker diarization of French broadcast news and talk shows on current affairs. This speaker diarization process is a multistage segmentation and clustering system. One of the stages is agglomerative clustering using state-of-the-art speaker identification methods (SID). For the QMMs used in this stage, we tried many different feature parameters, including MFCCs, Gaussianized MFCCs, Gaussianized MFCCs with cepstral mean subtraction, and Gaussianized MFCCs with cepstral mean substraction containing only frames with high energy. We found that this last set of feature parameters gave the best results. Compared to Gaussianized MFCCs, these features reduced the diarization error rate (DER) by 12% on a development set and by 19% on a test set. We also combined clusters resulting from Gaussianized and non-Gaussianized feature sets. This cluster combination resulted in another 4% reduction in DER for both the development and the test sets. The best DER we have achieved is 15.4% on the development set, and 14.5% on the test set.

It is not too difficult to select a fairly small (on the order of 20 000 words) fixed recognition vocabulary that will cover over 99% of new input words when the task is limited to text in a specific knowledge domain and when one disregards names and acronyms. Achieving such a level of coverage is much more difficult when restrictions on knowledge domain and names are lifted, however. This report describes how we selected a 75 000-word English recognition vocabulary that covers over 98% of words in new newspaper text, including names and acronyms. Observations collected during the vocabulary selection process indicate the limiting factors for coverage of general knowledge domain text such as newspaper stories.

This paper proposes a new way of using vector quantization for improving recognition performance for a 60,000 word vocabulary speaker-trained isolated word recognizer using a phonemic Markov model approach to speech recognition. We show that we can effectively increase the codebook size by dividing the feature vector into two vectors of lower dimensionality, and then quantizing and training each vector separately. For a small codebook size, integration of the results of the two parameter vectors provides significant improvement in recognition performance as compared to the quantizing and training of the entire feature set together. Even for a codebook size as small as 64, the results obtained when using the new quantization procedure are quite close to those obtained when using Gaussian distribution of the parameter vectors.

A novel type of hidden Markov model (HMM) has been developed to account explicitly for the context-dependent vowel acoustic transitions in consonant-vowel and vowel consonant phonetic environments. The major difference between this type of HMM and the standard Gaussian HMM is that the Gaussian mean vectors associated with the vowel HMM states, which are intended to model the vowel acoustic transitions, are set to be linearly interpolated values between those of the vowel steady state and those of the assumed locus for the adjacent consonant. The locus vectors, one for each consonant except for /h/, are trained together with all other HMM parameters using the Baum-Welch algorithm. The training procedure is fully automatic and converges to a local maximum. The model is incorporated in a phonetically based 75000-word vocabulary speech recognizer and provides a modest improvement in recognition rate over the standard approach. >

The acoustic recognizer of the INRS-Telecommunications 60000-word-vocabulary isolated-word recognition system is discussed. The task of the acoustic recognizer is to generate a list of word hypotheses and their likelihoods based on the acoustic data for each input word. Two sets of experiments are reported in which such knowledge is incorporated into the hidden Markov models (HMMs) used during recognition. In the first set, vowel duration properties are used in the HMMs. In the second set, word-initial and word-final stop consonants are modeled as a sequence of context-dependent subphonemes. The performance of the recognizer is significantly improved by appropriate utilization of the context-dependent vowel-duration information and the context-dependent microsegmental properties of stop consonants. >

Relative performance is compared for three different language models applied to the linguistic decoding part of a 75000-word speech recognizer. These models are the trigram model, the tri-POS model (POS stands for parts of speech), and a smoothed trigram model with tied distributions for words three or more syllables long. The full trigram model gives the best performance but is most expensive in terms of data and storage requirements. The smoothed trigram and tri-POS models yield equivalent performance. For general text entry tasks, use of the tri-POS model is suggested since it is less sensitive to variations in the discourse domains. For applications specific to individual discourse domains, trigram models trained on data obtained from the target domain are recommended. >

Recent results obtained with a hidden Markov model (HMM)-based acoustic recognizer using a virtually unlimited vocabulary (86000 words) to perform speaker-dependent isolated-word recognition are described. The task domain of this recognizer is quite general, consisting of paragraphs read from various newspapers, books, and magazines. The results of a comparative acoustic recognition study using various types of HMMs and various amounts of training data (from 700 to about 4000 words) are presented. The models explored include context-dependent allophonic HMMs (including generalized diphone and triphone models with unimodal Gaussian output densities) and context-independent phonemic HMMs (using either unimodal or mixture densities). Experimental results indicate that phonemic HMMs with many components in the mixture output densities provide the highest acoustic recognition accuracy. The acoustic recognition accuracy for a total of about 7000 test words spoken by four male and five female speakers is 82%. Recognition accuracy after application of the language model increases to 92%. >

In North America, people call the directory assistance operator to find the phone number of a business or residential listing. The directory assistance service is generally maintained by telcos, and it represents a significant cost to them. Partial or complete automation of directory assistance would result in significant cost savings for telcos. Nortel Networks has a product called Automated Directory Assistance System (ADAS) Plus which partially automates this directory assistance function through the use of speech recognition. The system has been deployed all across Quebec, through most of US West and BellSouth. ADAS Plus primarily automates the response to the question “for what city?” through speech recognition. We give details of this speech recognition system and outline its performance in the deployed regions.

This paper proposes simultaneous decoding using multiple utterances to derive one or more allophonic transcriptions for each word. Three possible simultaneous decoding algorithms, namely the N-best-based algorithm, the forward-backward-based algorithm and the word-network-based algorithm, are outlined. The proposed word-network-based algorithm can incrementally decode a transcription from any number of training utterances. Speech recognition experiments for both known and unknown word vocabularies show up to 16% reduction in word error rate when simultaneously decoded allophonic transcriptions are added to the recognition dictionaries. This result holds even for dictionaries originally transcribed by expert phoneticians.

Many applications require recognition of spoken isolated words or phrases from a large vocabulary. For example, the goal of the 86000-word recognizer at INRS-Telecommunications [14] is to transcribe speech spoken as a sequence of isolated words. The sentences to be read are chosen arbitrarily from a variety of sources, including newspapers, books, magazines, etc. Another example is the StockTalk system running at BNR Montreal [24] which dispenses real time stock quotes by voice over the telephone for stocks traded in New York, Toronto and NASDAQ stock exchanges. The vocabulary for this system consists of words or phrases spoken in isolation. This system requires speaker-independent recognition over the telephone, while the first example requires speaker-dependent recognition over high quality microphones.

The authors have assessed the possibility of modeling phone trajectories to accomplish speech recognition. This approach has been considered as one of the ways to model context-dependency in speech recognition based on the acoustic variability of phones in the current database. A hybrid segmental learning vector quantization/hidden Markov model (SLVQ/HMM) system has been developed and evaluated on a telephone speech database. The authors have obtained 85.27% correct phrase recognition with SLVQ alone. By combining the likelihoods issued by SLVQ and by HMM, the authors have obtained 94.5% correct phrase recognition, a small improvement over that obtained with HMM alone. >

We have developed an algorithm for the automatic conversion of dictated English sentences to written text, with essentially no restriction on the nature of the material dictated. We require that speakers undergo a short training session so that the system can adapt to their individual speaking characteristics and that they leave brief pauses between words. We have tested our algorithm extensively on an 86,000 word vocabulary (the largest of any such system in the world) using nine speakers and obtained word recognition rates on the order of 93%.

We describe how we modified the CRIM factor analysis speaker verification system to handle the new cross-channel conditi ons encountered in the 2008 NIST speaker recognition evaluation. Using the 2006 evaluation data for development, we obtained results on a broad spectrum of test conditions that are uniformly bette r than the best results that have been published in the literature. Index Terms: speaker verification, factor analysis

We report results on speaker diarization of telephone conversations. This speaker diarization process is similar to the multistage segmentation and clustering system used in broadcast news. It consists of an initial acoustic change point detection algorithm, iterative Viterbi re-segmentation, gender labeling, agglomerative clustering using a Bayesian information criterion (BIC), followed by agglomerative clustering using state-of-the-art speaker identification methods (SID) and Viterbi re-segmentation using Gaussian mixture models (GMMs). The Viterbi re-segmentation using GMMs is new, and it reduces the diarization error rate (DER) by 10%. We repeat these multistage segmentation and clustering steps twice: once with MFCCs as feature parameters for the GMMs used in gender labeling, SID and Viterbi re-segmentation steps, and another time with Gaussianized MFCCs as feature parameters for the GMMs used in these three steps. The resulting clusters from the parallel runs are combined in a novel way that leads to a significant reduction in the DER. On a development set containing 30 telephone conversations, this combination step reduced the DER by 20%. On another test set containing 30 telephone conversations, this step reduced the DER by 13%. The best error rate we have achieved is 6.7% on the development set, and 9.0% on the test set.

Virtually all recognition systems, both research and commercial, place severe limits on the vocabulary and syntactic structures a speaker may use. This recognition system is the most advanced in the world in terms of freedom from this constraint—it allows an essentially unrestricted English vocabulary and sentence structure—although it is currently required that each word be followed by a short pause. Many speech recognition systems make little use of syntactic and semantic information. A trigram language model is employed here that examines the statistical likelihood of each vocabulary word's occurrence as a function of the two preceding words. A global context‐sensitive parser is also used which accepts, as input, a long sequence of word candidates. (The speaker is not required to pronounce punctuation marks.) Using rules of English syntax, the parser identifies many erroneous candidates and estimates locations to insert punctuation in the output text. The combined parser and trigram model have raised t...