当前位置：搜档网 › Speech-based Annotation of Heterogeneous Multimedia Content Using Automatic Speech Recognit

Speech-based Annotation of Heterogeneous Multimedia Content Using Automatic Speech Recognit

Speech-based Annotation of Heterogeneous Multimedia Content Using Automatic Speech Recognition

CTIT-technical Report,version1.0,May2007

Marijn Huijbregts,Roeland Ordelman,Franciska de Jong

Human Media Interaction,University of Twente

Enschede,The Netherlands http://hmi.ewi.utwente.nl

Abstract

This paper reports on the setup and evaluation of robust speech recognition system parts, geared towards transcript generation for heterogeneous,real-life media collections.The system

is deployed for generating speech transcripts for the NIST/TRECVID-2007test collection,

part of a Dutch real-life archive of news-related genres.Performance?gures for this type of

content are compared to?gures for broadcast news test data.

1Introduction

The exploitation of linguistic content such as transcripts generated via automatic speech recogni-tion(ASR)can boost the accessibility of multimedia archives enormously.This e?ect is of course limited to video data containing textual and/or spoken content but when available,the exploita-tion of linguistic content for the generation of a time-coded index can help to bridge the semantic gap between media features and search needs.This is con?rmed by the results of TREC series of Workshops on Video Retrieval(TRECVID)1.The TRECVID test collections contain not just video,but also ASR-generated transcripts of segments containing speech.Systems that do not exploit these transcripts typically do not perform as well as the systems that do incorporate speech features in their models[13],or to video content with links to related textual documents,such as subtitles and generated transcripts.

ASR supports the conceptual querying of video content and the synchronization to any kind of textual resource that is accessible,including other full-text annotation for audiovisual material[4]. The potential of ASR-based indexing has been demonstrated most successfully in the broadcast news domain.Spoken document retrieval in the American-English broadcast news(BN)domain was even declared‘a solved problem’based on the results of the TREC Spoken Document Retrieval (SDR)track in1999[7].Partly because collecting data to train recognition models for the BN domain is relatively easy,word-error-rates(WER)below10%are no longer exceptional[8,9],and ASR transcripts for BN content approximate the quality of manual transcripts,at least for several languages.

In other domains than broadcast news and for many less favored languages,a similar recogni-tion performance is usually harder to obtain due to(i)lack of domain-speci?c training data,and (ii)large variability in audio quality,speech characteristics and topics being addressed.However, as ASR performance of50%WER is regarded as a lower bound for successful retrieval,speech-based indexing for harder data remains feasible as long as the ASR performance is not below50% WER,and is actually a crucial enabling technology if no other means(metadata)are available to guide searching.

For2007,the TRECVID organisers have decided to shift the focus from broadcast news video to video from a real-life archive of news-related genres such as news magazine,educational,and 1https://www.sodocs.net/doc/7c16623775.html,

cultural programming.As in previous years,ASR transcripts of the data are provided as an optional information source for indexing.Apart from some English BN rushes(raw footage), the2007TRECVID collection consists of400hours of Dutch news magazine,science news,news reports,documentaries,educational programmes and archival video.The?les were provided by the Netherlands Institute for Sound and Vision2.(In the remainder this collection will be referred to as Sound and Vision data.)

This paper reports on the setup and evaluation of the speech recognition system(further referred to as SHoUT system3that is deployed for generating the transcripts that via NIST will be made available to the TRECVID-2007participants.The SHoUT system is particularly geared towards transcript generation for the kind of heterogeneous,real-life media collections exempli?ed by the Sound and Vision data that will feature in the TRECVID2007collection.In other words, it targets adequate retrieval performance,rather than plain robust ASR.

As can be expected for a diverse content set such as the Sound and Vision data,the audio and speech conditions vary enormously,ranging from read speech in a studio environment to spon-taneous speech under degraded acoustic conditions.Furthermore,a large variety of topics are addresses and the material dates from a broad time period.Historical items as well as contem-porary video fall within the range.(The former with poorly preserved audio;latter with varying audio characteristics,some even without’intended’sound,just noise).

To reach recognition accuracy that is acceptable for retrieval given the di?cult conditions,the di?erent parts of our ASR system must be made as robust as possible so that it is able to cope with those problems(such as mismatches between training and testing conditions)that typically emerge when technology is transferred from the lab and applied in a real life context.This is in accordance with our overall research agenda:the development of robust ASR technology that can be ported to di?erent topic domains with a minimum e?ort.Only technology that complies with this last requirement can be successfully deployed for the wide range of spoken word archives that call for annotation based on speech content such as cultural heritage data(historical archives and interviews),lecture collections,meeting archives(city council sessions,corporate meetings).

The remainder of this paper is organised as follows.Section2provides an overview of the system structure,followed by a description of the training procedure that was applied for the SHoUT-2007a version aiming at the transcription of Sound and Vision data(section3).Section4 presents and discusses performance evaluation results of this system obtained using Sound and Vision data and various other available test corpora.

2System structure

Figure1is a graphical representation of the system work?ow that starts with speech activity detection(SAD)in order to?lter out the audio that do not contain speech.The audio in the Sound and Vision collection contains all kinds of sounds such as music,sound e?ects or background noise with high volume(tra?c,cheering audience,etc).As a speech decoder will always try to map a sound segment to a sequence of words,processing non-speech portions of the videos(i) would be a waste of processor time,(ii)introduces noise in the transcripts due to assigning word labels to non-speech fragments,and(iii)reduces speech recognition accuracy in general when the output of the?rst recognition run is used for acoustic model adaptation purposes.

Figure1:Overview of the decoding system.Each step provides input for the following step.

2Netherlands Institute for Sound and Vision:http://www.beeldengeluid.nl/

3SHoUT is an acronym for SpeecH recognition University of Twente.

After SAD,the speech fragments are segmented and clustered.In this step,the speech frag-ments are split into segments that only contain speech from one single speaker.Each segment is labelled with its corresponding speaker ID.Next,for each segment the vocal tract length(VTLN) warping factor is determined for vocal tract length normalisation during feature extraction for the?rst decoding step.Decoding is done using the HMM-based Viterbi decoder.In the?rst decoding iteration,triphone VTLN acoustic models and trigram language models are used.For each speaker,a?rst best hypothesis aligned on a phone basis is created for unsupervised acous-tic model adaptation.For each?le,the language model is mixed with a topic speci?c language model.The second decoding iteration uses the speaker adapted acoustic models and the topic speci?c language models to create the?nal?rst best hypothesis aligned on word basis.Also,for each segment,a word lattice is created.

The following sections describe each of the system parts in more detail.

2.1Speech activity detection

Most SAD systems are based on a Hidden Markov Model(HMM)with a number of parallel states. Each state contains a Gaussian Mixture Model(GMM)that is trained on a class of sounds such as speech,silence or music.The classi?cation is done by performing a Viterbi search using this HMM.

A disadvantage of this approach is that the GMMs need to be trained on data that matches the evaluation data.The performance of the system can drop signi?cantly if this is not the case. Because of the large variety in the Sound and Vision collection it is di?cult to determine a good set of data on which to train the silence and speech models.Also it is di?cult to determine what kind of extra models are needed to?lter out unknown audio fragments like music or sound e?ects and to?nd training data for those models.

The SAD system proposed by[1]at RT06s uses regions in the audio that have low or high energy levels to train new speech and silence models on the data that is being processed.The major advantage of this approach is that no earlier trained models are needed and therefore the system is robust for domain changes.We extended this approach for audio that contains fragments with high energy levels that are not speech.

Instead of using energy as an initial con?dence measure,our system uses the output of our broadcast news SAD system.This system is only trained on silence and speech,but because energy is not used as a feature and most non-speech data will?t the more general silence model better than the speech model,most non-speech will be classi?ed as silence.After the initial segmentation the data classi?ed as silence is used to train a new silence model and a sound model.The silence model is trained on data with low energy levels and the sound model on data with high energy levels.After a number of training iterations,the speech model is also re-trained.The result is an HMM based SAD system with three models(speech,non-speech and silence)that are trained solely on the data under evaluation.

2.2Segmentation and clustering

Similar to the SAD system,the segmentation and clustering system uses GMMs to model di?erent classes.This time instead of trying to distinguish speech from non-speech,each GMM represents a single unique speaker.In order to?nd the correct number of speakers,?rst the data is divided into a number of small segments and for each segment a model is trained.After this,all models are pairwise compared.If two models are very similar(if it is believed that they model the same speaker)they are merged.This procedure is repeated until no two models can be found that are believed to be trained on speech of the same speaker.In[14]this speaker diarization algorithm is discussed in depth.

Although this speaker diarization approach has very good clustering results,it is not very fast on long audio?les.In order to be able to process the entire Sound and Vision collection in reasonable time,we have changed the system slightly.For the purpose it is used here(to segment

the data so that we can apply VTLN and do unsupervised adaptation)a less accurate clustering will still be very helpful.

The majority of processing time is spent on pairwise comparing models.At each merging iteration,all possible model combinations need to be recomputed.The longer a?le is,the more initial clusters are needed.With this,the number of model combinations that need to be consid-ered for merging increases more than linearly.We managed to bring back the number of merge calculations by simply merging multiple models at a time.The two models A and B with the highest BIC score are merged?rst,followed by the two models C and D with the second highest score.If this second merge involves one of the earlier merged models,for example model C is the same model as model A,also the other combination(B,D)must have a positive BIC score.This process is repeated with a maximum of four merges at one iteration.Without performance loss (see section4),this procedure reduces the real-time factor from8times to2.7times real-time on a35minute TRECVID?le.

2.3Vocal tract length normalization

Variation of vocal tract length between speakers makes it harder to train robust acoustic models. In the SHoUT system,normalisation of the feature vectors is obtained by shifting the Mel-scale windows by a certain warping factor.

If the warping factor is smaller than one,the windows will be stretched and if the factor is bigger than one,the windows will be compressed.To normalize for the vocal tract length,large warping factors should be applied for people with a low voice and small warping factors for people with a high voice(note that this is typically because of how we implemented the warping factor. In the literature often big warping factors are linked to high voices instead of low voices).

In order to determine the speakers warping factors,a gaussian mixture model(referred to as VTLN-GMM)was trained with data from the Spoken Dutch Corpus(CGN)[10].In total,speech of200male speakers and200female speakers was used.The GMM only contained four gaussians. For the400speakers in the training set,the warping factor is determined by calculating the feature vectors with a warping factor varying from0.8to1.2in step sizes of0.04and determining for each of these feature sets what the probability on the VTLN-GMM is.For each speaker the warping factor used to create the set of features with the highest probability is chosen.

After each speaker is assigned a warping factor,a new VTLN-GMM is trained using normalised feature vectors and again the warping factors are determined by looking at a range of factors and picking the one with the highest score.This procedure is repeated a number of times so that a VTL normalised speech model is created.From this point on,the warping factor for each speaker is determined by looking at a range of factors on this normalised model.The acoustic models needed for decoding are trained on features that are normalised using this method.

Figure2contains the warping factors of the hundred speakers of the VTLN-GMM training set split in male and female.Three training iterations were performed.At the third iteration a model with eight gaussians was trained.This model is picked to be the?nal VTLN-GMM.

2.4Decoding

The decoder’s Viterbi search is implemented using the token passing paradigm[6].HMMs with three states and GMMs for its probability density functions are used to calculate acoustical like-lihoods of context dependent phones.Up to4-gram back-o?language models(LMs)are used to calculate the priors.

The HMMs are organised in a single Pronunciation Pre?x Tree(PPT)as described in[6,5]. This PPT contains the pronunciation of every word from the vocabulary.Each word is de?ned by an ID of four bytes.Identical words with di?erent pronunciations will receive the same ID making it possible to model pronunciation variations.

Instead of copying PPTs for each possible linguistic state(the LM n-gram history),each token contains a pointer to its LM history(see?gure3).Tokens coming from leaves of the PPT are fed back into the root node of the tree after their n-gram history is updated.Only for tokens with

Figure2:The warping factor of200male and200female voices determined using the VTLN-GMM after each of its four training iterations.The female speakers are clustered at the left and the male speakers at the right.

the same LM history,token collisions will occur.This means that each HMM state of each node in the single PPT can contain a list of tokens with unique n-gram histories.These lists are sorted in descending order of the token probability scores.

Figure3:The decoder uses a single Pronunciation Pre?x Tree(PPT)and a4-gram Language Model(LM).Each HMM state from the PPT can contain a sorted list of tokens.Each token from a list has a unique LM history.

The LM data are stored in up to four lookup tables.The?rst table contains unigram prob-abilities and back-o?values for all words of the lexicon.The statistics for all available bigrams, trigrams and fourgrams are stored in three minimal perfect hash tables[2].These hash tables con-tain exactly one occurrence in each slot of the table.This means that no extra memory is needed except for storing the hash function and for the key of each data structure,the n-gram history. This key is needed because during lookup,the hash function will map queries for non-existing n-grams to random slots.By comparing the n-gram of the query to the n-gram of the found table slot,it can be determined if the search is successful.The algorithm proposed in[3,2]is used to generate the hash functions.Each n-gram table contains a backo?value so that when an n-gram probability does not exist in the hash table,the system can backo?and look up the(n-1)-gram probability of the last words of the token n-gram history.

3System training

3.1SAD model training

The GMMs used to initialise the SAD system(see section2.1)are trained on a small amount of Dutch broadcast news training data from the CGN corpus[10].Three and a half hours of speech and half an hour of silence from200male and200female speakers were used.The feature vectors consist of twelve MFCC components and the zero-crossing statistic.Each GMM(silence and speech)contains20gaussians.

3.2Acoustic model training

The acoustic models(AMs)for ASR are trained using features with twelve MFCC components. Energy is added as a thirteenth component and deltas and delta-deltas are calculated.Before calculating the MFCC features,the warping factor is determined as described in section2.3so that the Mel-scale windows can be shifted according to this factor.

A set of39triphones and one silence phone is used to model acoustics.The triphones are modelled using three HMM states.During training a decision tree is used to tie the triphone gaussians for each state.The number of triphone clusters(triphones that are mapped to the same gaussian mixture)is limited by the amount of available training data for each cluster and a ?xed maximum allowed number of clusters.After clustering the triphones,the models are trained starting with one gaussian each and iteratively increasing the number of gaussians until each cluster contains32gaussians.

In total the acoustic models were trained on79hours of broadcast news,interviews and live commentaries of sport events from the CGN corpus[10].During training1443triphone clusters were de?ned resulting in acoustic models with in total46176gaussians.

This data collection is the training set that is allowed to use in the primary condition of the Dutch ASR benchmark organized by the N-Best project4.

3.3Acoustic model adaptation

The clustering information obtained during speaker diarization(see section2.2)is used to create speaker dependent acoustic models.SMAPLR adaptation([12])is used to adapt the means of the acoustic model gaussians.Before adapatation starts,a tree structure is created for the adaptation data.The more data is available,the deeper the tree will be.The root of the tree contains all data and the leaves only small sets of phones.The data at each branch is used to perform Maximum a Posteriori(MAP)adaptation using output of its parent nodes as prior densities for the transformation matrices.This method prevents over-adapting models with small amounts of data but it also proves to adapt models very well if more data is available.Most important,hardly any tuning is needed using this method.Only the minimum amount of data per tree node and the maximum tree depth needed to be de?ned.These two parameters have been tuned on broadcast news data.

3.4Language model training

Broadcast news language models(LMs)were estimated from approximately500million words of normalised Dutch text data from various resources(See Table1).The larger part of the text data consists of written texts(98%),of which the majority is newspaper data5.A small portion of the available resources,some1million words,consists of speech transcripts,mostly derived from the various domains of the Spoken Dutch Corpus(CGN)[10]and some collected at our institute as part of the BN-speech component of the Twente News Corpus(TwNC-speech).

4http://speech.tm.tno.nl/n-best

5provided for research purposes by the Dutch Newspaper publisher PCM.

corpus description word count

Spoken Dutch Corpus various 6.5M

Twente News Corpus(speech)’01-’02BN transcripts112K

Dutch BN Autocues Corpus’98-’05BN teleprompter 5.7M

Twente News Corpus(written)’99-’06Dutch NP513M

Table1:Dutch text corpora and word count based on normalised text.

For the selection of the vocabulary words we used the14K most frequent words from the CGN corpus with a minimum word frequency of10,a selection of recent words from2006newspaper (until31/08/2006,see below)data and a selection of most frequent words from the1999-2005 newspaper(NP)data set.The aim was to end up with a vocabulary of around50K words that could be regarded as a more or less stable,general BN vocabulary that could later be expanded using task-speci?c words learned from either meta-data or via multi-pass speech recognition ap-proaches.Table2shows that a selection of words from speech transcripts(CGN)augmented with recent newspaper data gave the best result with respect to OOV(2.33%).

vocabulary%OOV

14.6K CGN9.04

50K general NP(99-05) 2.88

50K recent NP 2.39

51K merge recent-NP,CGN 2.33

Table2:OOV rates using di?erent vocabulary selection strategies based on available training data.

Trigram language models were estimated from the available data sources using modi?ed Kneser-Ney smoothing.Perplexities of the LMs were computed using a set of manually transcribed BN shows of7K words that date from the period after the most recent data in our text data collection. From the best performing LMs interpolated versions were created.In Table3perplexity results and mixture-weights of the respective models are listed.The mixture-LM with the lowest perplexity (211)was was used for decoding and contained about8.9M2-grams and30M3-grams.

ID LM PP mixture-weigths

01Autocues Corpus LM389n.a.

02CGN comp-k(news)797n.a.

0301+TwNC-speech+02362n.a.

04Twente News Corpus(NP)266n.a.

05cgn-comp-f(discussion/debates)652n.a.

06mix03-042260.44-0.56

07mix03-04-052110.27-0.55-0.18

Table3:Language model perplexities

3.5Video item speci?c language models

For every video item in the Sound and Vision collection,descriptive metadata is provided contain-ing,among others,content summaries of around100words.In order to minimise the OOV rate for content words in the video items,item-speci?c language models were created using these descrip-tions,a database of newspaper articles(Twente News Corpus)and an Information Retrieval(IR) system that returns ranked lists of relevant documents given a query or‘topic’.The procedure was as follows:

https://www.sodocs.net/doc/7c16623775.html,e description of video item from metadata as a query

2.generate LM training data from top N most relevant documens given a query

3.select most frequent words

4.create topic speci?c language model

5.mix speci?c model with background language model using a?xed weight

The meta-data descriptions were normalised and stop-words were?ltered out in the query processing part of the IR.For the initial experiments described in this paper,the top50,250and 1000documents from the ranked lists were taken as input for language model training.Every word from the meta-data description was added to the new vocabulary.New words from the newspaper data were only included when they exceeded a minimum frequency count(10in the experiments reported here).Pronunciations for the new words were generated using an automatic grapheme-to-phoneme converter[11].As no example data was available for every video item to estimate mixture weights appropriately,a?xed weight was chosen(0.8for the NP-LM in the experiments reported here).

4Experimental results

4.1Evaluation data

For evaluation purposes we used a number of di?erent resources.For evaluation of SAD and speaker diarization we used RT06s and RT07s conference meetings.These data do not match the training conditions as will also be the case with the Sound and Vision data and should provide a nice comparison.Also,from the Sound and Vision data,13fragments of5minutes each were randomly selected and annotated manually.To compare speech recognition results on Sound and Vision data with broadcast news transcription performance we selected one recent broadcast news show from the Twente News Corpus(TwNC-BN)that dates after the most recent date of the data that was used for language model training.For global comparison we also selected broadcast news data from the Spoken Dutch Corpus(CGN-BN).

4.2Speech activity detection

In Table4,the SAD error rate on RT06s and RT07s meeting data and Sound and Vision data are listed.On the RT06s and Sound and Vision data we also evaluated our BN SAD system(baseline). The results show that on both evaluation corpora we improved on the baseline BN SAD system. The SAD error on the conference meeting data is conform state-of-the-art6.

eval baseline SHoUT-2007a

RT06s26.9%4.4%

RT07s3.3%

Sound and Vision18.3%10.4%

Table4:Evaluation results of the SAD component showing results of the BN system(baseline) and optimised results(SHoUT-2007a).

6Because of the rules of the NIST evaluation we are prohibited from comparing our results with those of other participants.See https://www.sodocs.net/doc/7c16623775.html,/speech/tests/rt/rt2006/spring/pdfs/rt06s-SPKR-SAD-results-v5.pdf for the actual ranking.

4.3Speaker diarization

As we did not have ground truth transcriptions on Sound and Vision or BN data for diarization, we could only test diarization on conference meeting data.The speaker diarization error(DER) on the conference meetings of RT07s is11.14%on the Multiple Distance Microphone(MDM)task. The DER on the Single Distance Microphone(SDM)task,where only one single microphone may be used which is more comparible to the task in our system,is17.28%.Both results are conform state-of-the-art6.

4.4Automatic speech recognition

Table5shows the evaluation results on automatic speech recognition.It contains the Word Error Rates(WER)of the systems with and without VTLN applied.For all conditions we see a stable improvement over the baseline.It can be observed that there is a substantial performance di?erence between the TwNC-BN set and the CGN-BN set.This may be because the TwNC-BN data is more di?cult or that the audio conditions in the CGN-BN data set better match the training conditions as a large part of our acoustic training material is derived from this corpus. Note that the CGN-BN data we used for evaluation were not in our acoustic model and language model training sets.Again it shows that Sound and Vision data is a di?cult task domain.

eval baseline SHoUT-2007a

TwNC-BN32.8%28.5%

CGN-BN22.1%19%

Sound and Vision68.4%64.0%

Table5:Evaluation results of the ASR component on di?erent evaluation corpora(eval)showing Word Error Rate without VTL normalization(baseline)and with normalization(SHoUT-2007a).

4.5Video item-speci?c language models

When we compared the baseline ASR results with the ASR approach that uses video item-speci?c LMs we observed that for the condition that uses the1000highest ranked documents for estimating the item-speci?c LMs some video items WER improved signi?cantly(up to3%absolute)whereas for others the improvement was only marginal or even negative.On average WER improved with 0.8%absolute.

item base item-LM delta

0173.773.5-0.2

0255.956.3+0.4

0353.750.7-3.0

0474.773.8-0.9

0579.279.4+0.2

0642.643.2+0.6

0760.357.9-2.4

0862.864.1+1.3

0959.056.6-2.4

1039.138.7-0.4

1174.172.4-1.7

all61.460.6-0.8

Table6:Results of baseline speech recognition on Sound and Vision data(base)compared with speech recognition results using a video item speci?c language models(item-LM).

In Table7baseline ASR results are compared with the ASR approach that uses video item speci?c LMs.Only the results for the condition that uses the1000highest ranked documents for estimating the item speci?c LMs are shown as this condition produced the best results.The results indicate that using item speci?c LMs generally helps to improve ASR performance.

item base item-LM delta

0173.773.5-0.2

0255.956.3+0.4

0353.750.7-3.0

0474.773.8-0.9

0579.279.4+0.2

0642.643.2+0.6

0760.357.9-2.4

0862.864.1+1.3

0959.056.6-2.4

1039.138.7-0.4

1174.172.4-1.7

all61.460.6-0.8

Table7:Results of baseline speech recognition on Sound and Vision data(base)compared with speech recognition results using a video item speci?c language models(item-LM).

5Conclusions

It is clear that ASR results on the Sound and Vision data leave room for improvement.When we average the results on the two types of BN data we end up with a BN-WER of23.7%so there is a large gap between performance on BN data and performance on our target domain. The assumption is that an important part of the error is due to mismatch between training and testing conditions(speakers,recording set-ups).By implementing noise reduction techniques such as successfully applied in[1],we expect to improve system robustness on this part.On the LM level,we will?ne-tune the video-speci?c LM algorithm and work on the lattice output of the system(rescoring with higher order n-grams).

6Acknowledgements

The work reported here was partly supported by the Dutch bsik-programme MultimediaN(http: //www.multimedian.nl)and the EU projects MESH(IST-FP6-027685),and MediaCampaign (IST-PF6-027413).We would like to thank all those who helped us with generating speech tran-scripts for the Sound and Vision test set.

References

[1]X.Anguera,C.Wooters,and J.Pardo.Robust speaker diarization for meetings:Icsi rt06s

evaluation system.In NIST Rich Transcription2006Spring Meeting Recognition Evaluation, RT06s,Washington DC,USA,volume4299of Lecture Notes in Computer Science,Berlin, October2007.Springer Verlag.

[2]A.Cardenal,J.Dieguez,and C.Garcia-Mateo.Fast lm look-ahead for large vocabulary

continuous speech recognition using perfect hashing.In proceedings ICASSP2002,pages 705–708,Orlando,USA,2002.

[3]Zbigniew J.Czech,George Havas,and Bohdan S.Majewski.An optimal algorithm for gen-

erating minimal perfect hash https://www.sodocs.net/doc/7c16623775.html,rmation Processing Letters,43(5):257–264,1992.

[4]F.M.G.de Jong,R.J.F.Ordelman,and M.A.H.Huijbregts.Automated speech and audio

analysis for semantic access to multimedia.In Y.Avrithis,Y.Kompatsiaris,S.Staab,and N.E.O’Connor,editors,Proceedings of the First International Conference on Semantic and Digital Media Technologies,SAMT2006,Athens,Greece,volume4306of Lecture Notes in Computer Science,pages226–240,Berlin,December2006.Springer Verlag.

[5]Kris Demuynck,Jacques Duchateau,Dirk Van Compernolle,and Patrick Wambacq.An

e?cient search space representation for large vocabulary continuous speech recognition.Speech Commun.,30(1):37–53,2000.

[6]M.Finke,J.Fritsch, D.Koll,and A.Waibel.Modeling and e?cient decoding of large

vocabulary conversational speech.In proceedings Eurospeech’99,pages467–470,Budapest, Hungary,1999.

[7]J.S.Garofolo,C.G.P.Auzanne,and E.M Voorhees.The TREC SDR Track:A Success Story.

In Eighth Text Retrieval Conference,pages107–129,Washington,2000.

[8]Jean-Luc Gauvain,Gilles Adda,Martine Adda-Decker,Alexandre Allauzen,Veronique Gend-

ner,Lori Lamel,and Holger Schwenk.Where Are We in Transcribing French Broadcast News?

In InterSpeech,Lisbon,September2005.

[9]L.Nguyen,S.Abdou,M.A?fy,J.Makhoul,S.Matsoukas,R.Schwartz,B.Xiang,https://www.sodocs.net/doc/7c16623775.html,mel,

J.L.Gauvain,G.Adda,H.Schwenk,and F.Lefevre.The2004BBN/LIMSI10xRT English Broadcast News Transcription System.In Proc.DARPA RT04,Palisades NY,November 2004.

[10]N.Oostdijk.The Spoken Dutch Corpus.Overview and?rst evaluation.In M.Gravilidou,

G.Carayannis,S.Markantonatou,S.Piperidis,and G.Stainhaouer,editors,Second Interna-

tional Conference on Language Resources and Evaluation,volume II,pages887–894,2000.

[11]Roeland Ordelman.Dutch Speech Recognition in Multimedia Information Retrieval.PhD

thesis,University of Twente,The Netherlands,October2003.

[12]O.Siohan,T.Myrvol,and C.Lee.Structural maximum a posteriori linear regression for

fast hmm adaptation.In In ISCA ITRW Automatic Speech Recognition:Challenges for the Millenium,pages120–127,2000.

[13]Alan F.Smeaton,Paul Over,and Wessel Kraaij.Evaluation campaigns and trecvid.In MIR

2006-8th ACM SIGMM International Workshop on Multimedia Information Retrieval,2006.

[14]D.A.van Leeuwen and M.A.H.Huijbregts.The ami speaker diarization system for nist

rt06s meeting data.In NIST Rich Transcription2006Spring Meeting Recognition Evaluation, RT06s,Washington DC,USA,volume4299of Lecture Notes in Computer Science,pages371–384,Berlin,October2007.Springer Verlag.

如何写先进个人事迹

如何写先进个人事迹篇一：如何写先进事迹材料如何写先进事迹材料一般有两种情况：一是先进个人，如先进工作者、优秀党员、劳动模范等；一是先进集体或先进单位，如先进党支部、先进车间或科室，抗洪抢险先进集体等。无论是先进个人还是先进集体，他们的先进事迹，内容各不相同，因此要整理材料，不可能固定一个模式。一般来说，可大体从以下方面进行整理。 (1)要拟定恰当的标题。先进事迹材料的标题，有两部分内容必不可少，一是要写明先进个人姓名和先进集体的名称，使人一眼便看出是哪个人或哪个集体、哪个单位的先进事迹。二是要概括标明先进事迹的主要内容或材料的用途。例如《王鬃同志端正党风的先进事迹》、《关于评选张鬃同志为全国新长征突击手的材料》、《关于评选鬃处党支部为省直机关先进党支部的材料》等。 (2)正文。正文的开头，要写明先进个人的简要情况，包括：姓名、性别、年龄、工作单位、职务、是否党团员等。此外，还要写明有关单位准备授予他(她)什么荣誉称号，或给予哪种形式的奖励。对先进集体、先进单位，要根据其先进事迹的主要内容，寥寥数语即应写明，不须用更多的文字。然后，要写先进人物或先进集体的主要事迹。这部分内容是全篇材料

的主体，要下功夫写好，关键是要写得既具体，又不繁琐；既概括，又不抽象；既生动形象，又很实在。总之，就是要写得很有说服力，让人一看便可得出够得上先进的结论。比如，写一位端正党风先进人物的事迹材料，就应当着重写这位同志在发扬党的优良传统和作风方面都有哪些突出的先进事迹，在同不正之风作斗争中有哪些突出的表现。又如，写一位搞改革的先进人物的事迹材料，就应当着力写这位同志是从哪些方面进行改革的，已经取得了哪些突出的成果，特别是改革前后的．经济效益或社会效益都有了哪些明显的变化。在写这些先进事迹时，无论是先进个人还是先进集体的，都应选取那些具有代表性的具体事实来说明。必要时还可运用一些数字，以增强先进事迹材料的说服力。为了使先进事迹的内容眉目清晰、更加条理化，在文字表述上还可分成若干自然段来写，特别是对那些涉及较多方面的先进事迹材料，采取这种写法尤为必要。如果将各方面内容材料都混在一起，是不易写明的。在分段写时，最好在每段之前根据内容标出小标题，或以明确的观点加以概括，使标题或观点与内容浑然一体。最后，是先进事迹材料的署名。一般说，整理先进个人和先进集体的材料，都是以本级组织或上级组织的名义；是代表组织意见的。因此，材料整理完后，应经有关领导同志审定，以相应一级组织正式署名上报。这类材料不宜以个人名义署名。写作典型经验材料-般包括以下几部分： (1)标题。有多种写法，通常是把典型经验高度集中地概括出来，一

Java注解

注解可以先把注解当成注释来看，注释就是给类的各个组成部分（包、类名、构造器、属性、方法、方法参数，以及局部变量）添加一些解释。可以先不去管注解是用来干什么的，就把它当成注释来看。注解的格式当然不能与注释相同，注解是需要声明的，声明注解与声明一个接口有些相似。当然Java也有一些内置注解，例如：@Override就是内置注解。 1声明注解声明注解与声明一个接口相似，它需要使用@interface。一个注解默认为Annotation的注解还可以带有成员，没有成员的注解叫做标记注解。成员的类型只能是基本类型、枚举类型）、String、基本类型数组、String[]，以及注解和注解数组类型。其中String表示成员的类型，value()表示成员名称。其中圆括号不能没有，也不能在圆

括号内放参数，它不是一个方法，只是一个成员变量。注解可以有多个成员，但如果只有一个成员，那么成员名必须为value。这时在设置成

Java还提供了一些元注解，用来控制注解，例如@Retention和@Target： ●@Target：ElementType类型（枚举类型），表示当前注解可以标记什么东西，可选值为： TYPE：可以标记类、接口、注解类、Enum。 FIELD：可以标记属性。 METHOD：可以标记就去。 PARAMETER：可以标记参数。 CONSTRUCTOR：可以标记构造器。 LOCAL_VARIABLE：可以标记局部变量。 ANNOTATION_TYPE：可以标记注解类声明。

PACKAGE：可以标记包。 ●@Retention：RetentionPolicy类型（枚举类型），表示注解的可保留期限。可选值为： SOURCE：只在源代码中存在，编译后的字节码文件中不保留注解信息。 CLASS：保留到字节码文件中，但类加载器不会加载注解信息到JVM。 RUNTIME：保留到字节码文件中，并在目标类被类加载器加载时，同时加载注解信息到JVM，可以通过反射来获取注解信息。 2访问注解很多第三方程序或工具都使用了注解完成特殊的任务，例如Spring、Struts等。它们都提供了自己的注解类库。在程序运行时使用反射来获取注解信息。下面我们来使用反射来获取注解信息。

annotation入门_

Java Annotation 入门
摘要：本文针对 java 初学者或者 annotation 初次使用者全面地说明了 annotation 的使用方法、定义方式、分类。初学者可以通过以上的说明制作简单的 annotation 程序，但是对于一些高级的 an notation 应用（例如使用自定义 annotation 生成 javabean 映射 xml 文件）还需要进一步的研究和探讨。涉及到深入 annotation 的内容，作者将在后文《Java Annotation 高级应用》中谈到。
同时，annotation 运行存在两种方式：运行时、编译时。上文中讨论的都是在运行时的 annota tion 应用，但在编译时的 annotation 应用还没有涉及，
一、为什么使用 Annotation：
在 JAVA 应用中，我们常遇到一些需要使用模版代码。例如，为了编写一个 JAX-RPC web serv ice，我们必须提供一对接口和实现作为模版代码。如果使用 annotation 对远程访问的方法代码进行修饰的话，这个模版就能够使用工具自动生成。另外，一些 API 需要使用与程序代码同时维护的附属文件。例如，JavaBeans 需要一个 BeanIn fo Class 与一个 Bean 同时使用/维护，而 EJB 则同样需要一个部署描述符。此时在程序中使用 a nnotation 来维护这些附属文件的信息将十分便利而且减少了错误。
二、Annotation 工作方式：
在 5.0 版之前的 Java 平台已经具有了一些 ad hoc annotation 机制。比如，使用 transient 修饰符来标识一个成员变量在序列化子系统中应被忽略。而@deprecated 这个 javadoc tag 也是一个 ad hoc annotation 用来说明一个方法已过时。从 Java5.0 版发布以来，5.0 平台提供了一个正式的 annotation 功能：允许开发者定义、使用自己的 annoatation 类型。此功能由一个定义 annotation 类型的语法和一个描述 annotation 声明的语法，读取 annotaion 的 API，一个使用 annotation 修饰的 class 文件，一个 annotation 处理工具（apt）组成。
1

关于时间管理的英语作文 manage time

How to manage time Time treats everyone fairly that we all have 24 hours per day. Some of us are capable to make good use of time while some find it hard to do so. Knowing how to manage them is essential in our life. Take myself as an example. When I was still a senior high student, I was fully occupied with my studies. Therefore, I hardly had spare time to have fun or develop my hobbies. But things were changed after I entered university. I got more free time than ever before. But ironically, I found it difficult to adjust this kind of brand-new school life and there was no such thing called time management on my mind. It was not until the second year that I realized I had wasted my whole year doing nothing. I could have taken up a Spanish course. I could have read ten books about the stories of successful people. I could have applied for a part-time job to earn some working experiences. B ut I didn’t spend my time on any of them. I felt guilty whenever I looked back to the moments that I just sat around doing nothing. It’s said that better late than never. At least I had the consciousness that I should stop wasting my time. Making up my mind is the first step for me to learn to manage my time. Next, I wrote a timetable, setting some targets that I had to finish each day. For instance, on Monday, I must read two pieces of news and review all the lessons that I have learnt on that day. By the way, the daily plan that I made was flexible. If there’s something unexpected that I had to finish first, I would reduce the time for resting or delay my target to the next day. Also, I would try to achieve those targets ahead of time that I planed so that I could reserve some more time to relax or do something out of my plan. At the beginning, it’s kind of difficult to s tick to the plan. But as time went by, having a plan for time in advance became a part of my life. At the same time, I gradually became a well-organized person. Now I’ve grasped the time management skill and I’m able to use my time efficiently.

shiro入门教程

一、介绍： shiro是apache提供的强大而灵活的开源安全框架，它主要用来处理身份认证，授权，企业会话管理和加密。 shiro功能：用户验证、用户执行访问权限控制、在任何环境下使用session API，如cs程序。可以使用多数据源如同时使用oracle、mysql。单点登录(sso)支持。remember me服务。详细介绍还请看官网的使用手册：https://www.sodocs.net/doc/7c16623775.html,/reference.html 与spring security区别，个人觉得二者的主要区别是： 1、shiro灵活性强，易学易扩展。同时，不仅可以在web中使用，可以工作在任务环境内中。 2、acegi灵活性较差，比较难懂，同时与spring整合性好。如果对权限要求比较高的项目，个人建议使用shiro，主要原因是可以很容易按业务需求进行扩展。附件是对与shiro集成的jar整合及源码。二、shiro与spring集成 shiro默认的配置，主要是加载ini文件进行初始化工作，具体配置，还请看官网的使用手册（https://www.sodocs.net/doc/7c16623775.html,/web.html）init文件不支持与spring的集成。此处主要是如何与spring及springmvc集成。 1、web.xml中配置shiro过滤器，web.xml中的配置类使用了spring的过滤代理类来完成。 Xml代码 2、在spring中的application.xml文件中添加shiro配置：

Java代码

anon org.apache.shiro.web.filter.authc.AnonymousFilter authc org.apache.shiro.web.filter.authc.FormAuthenticatio nFilter authcBasic org.apache.shiro.web.filter.authc.BasicHttpAuthenti cationFilter logout org.apache.shiro.web.filter.authc.LogoutFilter noSessionCrea tion org.apache.shiro.web.filter.session.NoSessionCreati onFilter perms org.apache.shiro.web.filter.authz.PermissionsAuthor izationFilter port org.apache.shiro.web.filter.authz.PortFilter rest org.apache.shiro.web.filter.authz.HttpMethodPermiss ionFilter roles org.apache.shiro.web.filter.authz.RolesAuthorizatio nFilter ssl org.apache.shiro.web.filter.authz.SslFilter user https://www.sodocs.net/doc/7c16623775.html,erFilter

RESTEasy入门经典

RESTEasy是JBoss的开源项目之一，是一个RESTful Web Services框架。RESTEasy的开发者Bill Burke同时也是JAX-RS的J2EE标准制定者之一。JAX-RS 是一个JCP制订的新标准，用于规范基于HTTP的RESTful Web Services的API。我们已经有SOAP了，为什么需要Restful WebServices？用Bill自己的话来说："如果是为了构建SOA应用，从技术选型的角度来讲，我相信REST比SOAP更具优势。开发人员会意识到使用传统方式有进行SOA架构有多复杂，更不用提使用这些做出来的接口了。这时他们就会发现Restful Web Services的光明之处。" 说了这么多，我们使用RESTEasy做一个项目玩玩看。首先创造一个maven1的web 项目 Java代码 1.mvn archetype:create -DgroupId=org.bluedash \ 2. 3.-DartifactId=try-resteasy -DarchetypeArtifactId=maven-archetype -webapp 准备工作完成后，我们就可以开始写代码了，假设我们要撰写一个处理客户信息的Web Service，它包含两个功能：一是添加用户信息；二是通过用户Id，获取某个用户的信息，而交互的方式是标准的WebService形式，数据交换格式为XML。假设一条用户包含两个属性：Id和用户名。那么我们设计交换的XML数据如下: Java代码 1. 2. 1 3. liweinan 4. 首先要做的就是把上述格式转换成XSD2，网上有在线工具可以帮助我们完成这一工作3，在此不详细展开。使用工具转换后，生成如下xsd文件： Java代码 1. 2. 4.

关于管理的英语演讲

1.How to build a business that lasts100years 0:11Imagine that you are a product designer.And you've designed a product,a new type of product,called the human immune system.You're pitching this product to a skeptical,strictly no-nonsense manager.Let's call him Bob.I think we all know at least one Bob,right?How would that go? 0:34Bob,I've got this incredible idea for a completely new type of personal health product.It's called the human immune system.I can see from your face that you're having some problems with this.Don't worry.I know it's very complicated.I don't want to take you through the gory details,I just want to tell you about some of the amazing features of this product.First of all,it cleverly uses redundancy by having millions of copies of each component--leukocytes,white blood cells--before they're actually needed,to create a massive buffer against the unexpected.And it cleverly leverages diversity by having not just leukocytes but B cells,T cells,natural killer cells,antibodies.The components don't really matter.The point is that together,this diversity of different approaches can cope with more or less anything that evolution has been able to throw up.And the design is completely modular.You have the surface barrier of the human skin,you have the very rapidly reacting innate immune system and then you have the highly targeted adaptive immune system.The point is,that if one system fails,another can take over,creating a virtually foolproof system. 1:54I can see I'm losing you,Bob,but stay with me,because here is the really killer feature.The product is completely adaptive.It's able to actually develop targeted antibodies to threats that it's never even met before.It actually also does this with incredible prudence,detecting and reacting to every tiny threat,and furthermore, remembering every previous threat,in case they are ever encountered again.What I'm pitching you today is actually not a stand-alone product.The product is embedded in the larger system of the human body,and it works in complete harmony with that system,to create this unprecedented level of biological protection.So Bob,just tell me honestly,what do you think of my product? 2:47And Bob may say something like,I sincerely appreciate the effort and passion that have gone into your presentation,blah blah blah-- 2:56(Laughter) 2:58But honestly,it's total nonsense.You seem to be saying that the key selling points of your product are that it is inefficient and complex.Didn't they teach you 80-20?And furthermore,you're saying that this product is siloed.It overreacts, makes things up as it goes along and is actually designed for somebody else's benefit. I'm sorry to break it to you,but I don't think this one is a winner.

个人先进事迹简介

个人先进事迹简介 01 在思想政治方面,xxxx同学积极向上,热爱祖国、热爱中国共产党,拥护中国共产党的领导.利用课余时间和党课机会认真学习政治理论,积极向党组织靠拢. 在学习上,xxxx同学认为只有把学习成绩确实提高才能为将来的实践打下扎实的基础,成为社会有用人才.学习努力、成绩优良. 在生活中,善于与人沟通,乐观向上,乐于助人.有健全的人格意识和良好的心理素质和从容、坦诚、乐观、快乐的生活态度,乐于帮助身边的同学,受到师生的好评. 02 xxx同学认真学习政治理论,积极上进,在校期间获得原院级三好生,和校级三好生,优秀团员称号,并获得三等奖学金. 在学习上遇到不理解的地方也常常向老师请教,还勇于向老师提出质疑.在完成自己学业的同时,能主动帮助其他同学解决学习上的难题,和其他同学共同探讨,共同进步. 在社会实践方面,xxxx同学参与了中国儿童文学精品“悦”读书系,插画绘制工作,xxxx同学在班中担任宣传委员,工作积极主动,认真负责,有较强的组织能力.能够在老师、班主任的指导下独立完成学院、班级布置的各项工作. 03 xxx同学在政治思想方面积极进取,严格要求自己.在学习方面刻苦努力,不断钻研,学习成绩优异,连续两年荣获国家励志奖学金；作

为一名学生干部,她总是充满激情的迎接并完成各项工作,荣获优秀团干部称号.在社会实践和志愿者活动中起到模范带头作用. 04 xxxx同学在思想方面,积极要求进步,为人诚实,尊敬师长.严格要求自己.在大一期间就积极参加了党课初、高级班的学习,拥护中国共产党的领导,并积极向党组织靠拢. 在工作上,作为班中的学习委员,对待工作兢兢业业、尽职尽责的完成班集体的各项工作任务.并在班级和系里能够起骨干带头作用.热心为同学服务,工作责任心强. 在学习上,学习目的明确、态度端正、刻苦努力,连续两学年在班级的综合测评排名中获得第1.并荣获院级二等奖学金、三好生、优秀班干部、优秀团员等奖项. 在社会实践方面,积极参加学校和班级组织的各项政治活动,并在志愿者活动中起到模范带头作用.积极锻炼身体.能够处理好学习与工作的关系,乐于助人,团结班中每一位同学,谦虚好学,受到师生的好评. 05 在思想方面,xxxx同学积极向上,热爱祖国、热爱中国共产党,拥护中国共产党的领导.作为一名共产党员时刻起到积极的带头作用,利用课余时间和党课机会认真学习政治理论. 在工作上,作为班中的团支部书记,xxxx同学积极策划组织各类团活动,具有良好的组织能力. 在学习上,xxxx同学学习努力、成绩优良、并热心帮助在学习上有困难的同学,连续两年获得二等奖学金. 在生活中,善于与人沟通,乐观向上,乐于助人.有健全的人格意识和良好的心理素质.

关系映射annotation

一对一(One-To-One) 使用@OneToOne注解建立实体Bean之间的一对一关联。一对一关联有三种情况：(1).关联的实体都共享同样的主键，(2).其中一个实体通过外键关联到另一个实体的主键(注意要模拟一对一关联必须在外键列上添加唯一约束)，(3).通过关联表来保存两个实体之间的连接关系(要模拟一对一关联必须在每一个外键上添加唯一约束)。 1.共享主键的一对一关联映射： @Entity @Table(name="Test_Body") public class Body { private Integer id; private Heart heart; @Id public Integer getId() { return id; } public void setId(Integer id) { this.id = id; } @OneToOne @PrimaryKeyJoinColumn public Heart getHeart() { return heart; }

public void setHeart(Heart heart) { this.heart = heart; } } @Entity @Table(name="Test_Heart") public class Heart { private Integer id; @Id public Integer getId() { return id; } public void setId(Integer id) { this.id = id; } } 通过@PrimaryKeyJoinColumn批注定义了一对一关联 2.使用外键进行实体一对一关联： @Entity @Table(name="Test_Trousers") public class Trousers { @Id public Integer id;

自我管理演讲稿英语翻译

尊敬的领导，老师，亲爱的同学们，大家好！我是5班的梁浩东。今天早上我坐车来学校的路上，我仔细观察了路上形形色色的人，有开着小车衣着精致的叔叔阿姨，有市场带着倦容的卖各种早点的阿姨，还有偶尔穿梭于人群中衣衫褴褛的乞丐。于是我问自己，十几年后我会成为怎样的自己，想成为社会成功人士还是碌碌无为的人呢，答案肯定是前者。那么十几年后我怎样才能如愿以偿呢，成为一个受人尊重，有价值的人呢？正如我今天演讲的题目是：自主管理。大家都知道爱玩是我们孩子的天性，学习也是我们的责任和义务。要怎样处理好这些矛盾，提高自主管理呢？首先，我们要有小主人翁思想，自己做自己的主人，要认识到我们学习，生活这一切都是我们自己走自己的人生路，并不是为了报答父母，更不是为了敷衍老师。我认为自主管理又可以理解为自我管理，在学习和生活中无处不在，比如通过老师，小组长来管理约束行为和同学们对自身行为的管理都属于自我管理。比如我们到一个旅游景点，看到一块大石头，有的同学特别兴奋，会想在上面刻上：某某某到此一游话。这时你就需要自我管理，你需要提醒自己，这样做会破坏景点，而且是一种素质低下的表现。你设想一下，如果别人家小孩去你家墙上乱涂乱画，你是何种感受。同样我们把自主管理放到学习上，在我们想偷懒，想逃避，想放弃的时候，我们可以通过自主管理来避免这些，通过他人或者自己的力量来完成。例如我会制定作息时间计划表，里面包括学习，运动，玩耍等内容的完成时间。那些学校学习尖子，他们学习好是智商高于我们吗，其实不然，在我所了解的哪些优秀的学霸传授经验里，就提到要能够自我管理，规范好学习时间的分分秒秒，只有辛勤的付出，才能取得优异成绩。在现实生活中，无数成功人士告诉我们自主管理的重要性。十几年后我想成为一位优秀的，为国家多做贡献的人。亲爱的同学们，你们们？让我们从现在开始重视和执行自主管理，十几年后成为那个你想成为的人。谢谢大家！

java中注解的几大作用

@SuppressWarnings("deprecation")//阻止警告 @HelloAnnotation("当为value属性时，可以省掉属性名和等于号。") public static void main(String[]args)throws Exception{ System.runFinalizersOnExit(true); if(AnnotationTest.class.isAnnotationPresent(HelloAnnotation.class)){ HelloAnnotation helloAnnotation= (HelloAnnotation)AnnotationTest.class.getAnnotation(HelloAnnotation.class); System.out.println("color():"+helloAnnotation.color()); System.out.println("value():"+helloAnnotation.value()); System.out.println("author():"+helloAnnotation.author()); System.out.println("arrayAttr():"+helloAnnotation.arrayAttr().length); System.out.println("annotationAttr():"+helloAnnotation.annotationAttr().value()); System.out.println("classType(): "+helloAnnotation.classType().newInstance().sayHello("hello,ketty")); } } @Deprecated//自定义：备注过时的方法信息 public static void sayHello(){ System.out.println("hello,world"); } }

ERDAS IMAGINE快速入门

实验一ERDAS IMAGINE快速入门一、背景知识 ERDAS IMAGINE是美国ERDAS公司开发的遥感图像处理系统，后来被Leica公司合并。它以其先进的图像处理技术，友好、灵活的用户界面和操作方式，面向广阔应用领域的产品模块，服务于不同层次用户的模型开发工具以及高度的RS/GIS（遥感图像处理和地理信息系统）集成功能，为遥感及相关应用领域的用户提供了内容丰富而功能强大的图像处理工具，代表了遥感图像处理系统未来的发展趋势。 ERDAS IMAGINE是以模块化的方式提供给用户的，可使用户根据自己的应用要求、资金情况合理地选择不同功能模块及其不同组合，对系统进行剪裁，充分利用软硬件资源，并最大限度地满足用户的专业应用要求，目前的最高版本为9.1。ERDAS IMAGINE面向不同需求的用户，对于系统的扩展功能采用开放的体系结构以IMAGINE Essentials、IMAGINE Advantage、IMAGINE Professional的形式为用户提供了低、中、高三档产品架构，并有丰富的功能扩展模块供用户选择，使产品模块的组合具有极大的灵活性。 ?IMAGINE Essentials级：是一个花费极少的，包括有制图和可视化核心功能的影像工具软件。该级功能的重点在于遥感图像的输入、输出与显示；图像库的建立与查询管理；专题制图；简单几何纠正与非监督分类等。 ?IMAGINE Advantage级：是建立在IMAGINE Essential级基础之上的，增加了更丰富的图像光栅GIS和单片航片正射校正等强大功能的软件。IMAGINE Advantag提供了灵活可靠的用于光栅分析，正射校正，地形编辑及先进的影像镶嵌工具。简而言之，IMAGINE Advantage是一个完整的图像地理信息系统（Imaging GIS）。 ?IMAGINE Professional级：是面向从事复杂分析，需要最新和最全面处理工具，

关于时间管理的英语演讲

Dear teacher and colleagues: my topic is on “spare time”. It is a huge blessing that we can work 996. Jack Ma said at an Ali's internal communication activity, That means we should work at 9am to 9pm, 6 days a week .I question the entire premise of this piece. but I'm always interested in hearing what successful and especially rich people come up with time .So I finally found out Jack Ma also had said :”i f you don’t put out more time and energy than others ,how can you achieve the success you want? If you do not do 996 when you are young ,when will you ?”I quite agree with the idea that young people should fight for success .But there are a lot of survival activities to do in a day ,I want to focus on how much time they take from us and what can we do with the rest of the time. As all we known ,There are 168 hours in a week .We sleep roughly seven-and-a-half and eight hours a day .so around 56 hours a week . maybe it is slightly different for someone . We do our personal things like eating and bathing and maybe looking after kids -about three hours a day .so around 21 hours a week .And if you are working a full time job ,so 40 hours a week , Oh! Maybe it is impossible for us at

优秀党务工作者事迹简介范文

优秀党务工作者事迹简介范文优秀党务工作者事迹简介范文 ***，男，198*年**月出生，200*年加入党组织，现为***支部书记。从事党务工作以来，兢兢业业、恪尽职守、辛勤工作，出色地完成了各项任务，在思想上、政治上同党中央保持高度一致，在业务上不断进取，团结同事，在工作岗位上取得了一定成绩。一、严于律己，勤于学习作为一名党务工作者，平时十分注重知识的更新，不断加强党的理论知识的学习，坚持把学习摆在重要位置，学习领会和及时掌握党和国家的路线、方针、政策，特别是党的十九大精神，注重政治理论水平的提高，具有坚定的理论信念;坚持党的基本路线，坚决执行党的各项方针政策，自觉履行党员义务，正确行使党员权利。平时注重加强业务和管理知识的学习，并运用到工作中去，不断提升自身工作能力，具有开拓创新精神，在思想上、政治上和行动上时刻同党中央保持高度一致。二、求真务实，开拓进取在工作中任劳任怨，踏实肯干，坚持原则，认真做好学院的党务工作，按照党章的要求，严格发展党员的每一个步骤，认真细致的对待每一份材料。配合党总支书记做好学院的党建工作，完善党总支建设方面的文件、材料和工作制度、管理制度等。

三、生活朴素，乐于助人平时重视与同事间的关系，主动与同事打成一片，善于发现他人的难处，及时妥善地给予帮助。在其它同志遇到困难时，积极主动伸出援助之手，尽自己最大努力帮助有需要的人。养成了批评与自我批评的优良作风，时常反省自己的工作，学习和生活。不但能够真诚的指出同事的缺点，也能够正确的对待他人的批评和意见。面对误解，总是一笑而过，不会因为误解和批评而耿耿于怀，而是诚恳的接受，从而不断的提高自己。在生活上勤俭节朴，不铺张浪费。身为一名老党员，我感到责任重大，应该做出表率，挤出更多的时间来投入到**党总支的工作中，不找借口，不讲条件，不畏困难，将总支建设摆在更重要的位置，解开工作中的思想疙瘩，为攻坚克难铺平道路，以支部为纽带，像战友一样团结，像家庭一样维系，像亲人一样关怀，践行入党誓言。把握机遇，迎接挑战，不负初心。

一小时搞明白注解处理器(Annotation Processor Tool)

一小时搞明白注解处理器（Annotation Processor Tool）什么是注解处理器？注解处理器是（Annotation Processor）是javac的一个工具，用来在编译时扫描和编译和处理注解（Annotation）。你可以自己定义注解和注解处理器去搞一些事情。一个注解处理器它以Java代码或者（编译过的字节码）作为输入，生成文件（通常是java文件）。这些生成的java文件不能修改，并且会同其手动编写的java代码一样会被javac编译。看到这里加上之前理解，应该明白大概的过程了，就是把标记了注解的类，变量等作为输入内容，经过注解处理器处理，生成想要生成的java代码。处理器AbstractProcessor 处理器的写法有固定的套路，继承AbstractProcessor。如下： [java] view plain copy 在CODE上查看代码片派生到我的代码片 public class MyProcessor extends AbstractProcessor { @Override public synchronized void init(ProcessingEnvironment processingEnv) { super.init(processingEnv); } @Override public Set getSupportedAnnotationTypes() { return null; } @Override public SourceVersion getSupportedSourceVersion() { return https://www.sodocs.net/doc/7c16623775.html,testSupported(); } @Override public boolean process(Set annotations, RoundEnvironment roundEnv) { return true; } } init(ProcessingEnvironment processingEnv) 被注解处理工具调用，参数ProcessingEnvironment 提供了Element，Filer，Messager等工具 getSupportedAnnotationTypes() 指定注解处理器是注册给那一个注解的，它是一个字符串的集合，意味着可以支持多个类型的注解，并且字符串是合法全名。getSupportedSourceVersion 指定Java版本 process(Set annotations, RoundEnvironment roundEnv) 这个也是最主