Flashcards
synthesis
Cards in this deck (40)
- 1formantan approach that synthesizes acoustics and formants based on rules and filters
- 2concatenative synthesisuse of databases of stored speech to assemble new utterance
- 3articulatory synthessimodeling of the movement of the articulatory and the acoustics of the vocal tract
- 4advtanges to formantvery intelligeble, avoids clipping artefacts between phonemes and is inexpensive
- 5disadvantagesunnatural robitic-sounding speech
- 6concat synhuman like natural sounding speech
- 7dis of concatto be robust needs a lot of data (comp expensive)
- 8advantages articulatoryinexpensive
- 9dis of articulatorynot very natural, it can be difficult to modify these systems to imitate new synthetic speaker
- 10homographsame spelling different meaning
- 11letter to soundsneed to find alignments, learn learn them through maximum likelihood estimation.
If you dont have access do expectation-maximization statistical machine translation. - 12induction of letter to sound rulesalignments can be improved by using hand-written rules that restruct the translation of letters to phonemes.
- 13Three aspects of prosody in TTSprominence, structure, tune
- 14prominencesome syllables or words are more prominent than others, especially content words
- 15structuresentences have inherent prosodic structure. Some word groups naturally join together, other require noticeble disjunction
- 16tuneto sound natural one has to account for the melody of an utterance
- 17diphonemiddle of a phonemen to the middle of the next
- 18duration modificationduplication or removable of short term periodic sequences
- 19pitchduration modification can be as simple as squishing or stresshing signals using decimation ro interpolation.
- 20bibliographic couplingmeasures of similarity between documents according to the intersection of their citations.
- 21co-citationmeasures of similarity between documents according to teh intersection of the documents that cite them
- 22vector space modelqueries and documents are both presented by unit length vectors in word space
- 23tfijterm frequency of word i in document j
- 24damped weight1+logtfij if tfij >0
0 if tfij = 0 - 25cosine methoddetermines the similarity between a query and some document
- 26dfinumber of documents in which wi appears
- 27dfimeaningful words may occour repeatedly in related documents but functional words may be distributed evenly over all documents
- 28inverse document frequencylog(D/dfi)
where D is total number of documents and we scale with log. - 29tf.idf(wi,dj)`(1+log(tfij)*log(D/dfi))
- 30co-occurrencewhen two or more terms occur in the same document more often than by chance.
- 31Latent semantic indexingproject qs and ds to a space with latent semantic dimensions.
- 32SVDsingular value decomposition
- 33A = T_(t*n)*S_n*n*(D_d*n)Tn =min(t,d)
- 34T and D are orthonormalall columns are orthogonal to each other.
- 35Naive BayesImagine each of D documents is a class with one training example- the document itself.
- 36Method for naive bayes rankingrank documents dj based on the posterior probabilty of a q being generated by those documents.
P(q| dj) - 37Linear interpolationlambda(P(q;dj))+1-lamdaPq
- 38precisionproportion of returned documents that are relevant
- 39recallproportion of relevant documents that are returned.
- 40extractive summarization
Information
- Description
- Flashcards on synthesis, created by arturo mena on 28/04/2017.
- Created by
- arturo mena
- Views
- 0
- Rating