vowel (Q6033078)

**Author**: Peter Turney (peter@ai.iit.nrc.ca) \N**Source**: [UCI](https://archive.ics.uci.edu/ml/machine-learning-databases/undocumented/connectionist-bench/vowel/) - date unknown \N**Please cite**: [UCI citation policy](https://archive.ics.uci.edu/ml/citation_policy.html)\N\N**Vowel Recognition (Deterding data)**\NSpeaker independent recognition of the eleven steady state vowels of British English using a specified training set of lpc derived log area ratios.\NCollected by David Deterding (data and non-connectionist analysis), Mahesan Niranjan (first connectionist analysis), Tony Robinson (description, program, data, and results)\N\NA very comprehensive description including comments by the authors can be found [here](https://archive.ics.uci.edu/ml/machine-learning-databases/undocumented/connectionist-bench/vowel/vowel.names)\N\NThe problem is specified by the accompanying data file, "vowel.data". This\Nconsists of a three dimensional array: voweldata [speaker, vowel, input].\NThe speakers are indexed by integers 0-89. (Actually, there are fifteen\Nindividual speakers, each saying each vowel six times.) The vowels are\Nindexed by integers 0-10. For each utterance, there are ten floating-point\Ninput values, with array indices 0-9.\N\NThe problem is to train the network as well as possible using only on data\Nfrom "speakers" 0-47, and then to test the network on speakers 48-89,\Nreporting the number of correct classifications in the test set.\N\NFor a more detailed explanation of the problem, see the excerpt from Tony\NRobinson's Ph.D. thesis in the COMMENTS section. In Robinson's opinion,\Nconnectionist problems fall into two classes, the possible and the\Nimpossible. He is interested in the latter, by which he means problems\Nthat have no exact solution. Thus the problem here is not to see how fast\Na network can be trained (although this is important), but to maximise a\Nless than perfect performance.\N\N#### METHODOLOGY\N\NReport the number of test vowels classified correctly, (i.e. the number of\Noccurences when distance of the correct output to the actual output was the\Nsmallest of the set of distances from the actual output to all possible\Ntarget outputs).\N\NThough this is not the focus of Robinson's study, it would also be useful\Nto report how long the training took (measured in pattern presentations or\Nwith a rough count of floating-point operations required) and what level of\Nsuccess was achieved on the training and testing data after various amounts\Nof training. Of course, the network topology and algorithm used should be\Nprecisely described as well.\N\N#### VARIATIONS\N\NThis benchmark is proposed to encourage the exploration of different node\Ntypes. Please theorise/experiment/hack. The author (Robinson) will try to\Ncorrespond by email if requested. In particular there has been some\Ndiscussion recently on the use of a cross-entropy distance measure, and it\Nwould be interesting to see results for that.\N\N#### Notes\N\N1. Each of these numbers is based on a single trial with random starting\Nweights. More trials would of course be preferable, but the computational\Nfacilities available to Robinson were limited.\N\N2. Graphs are given in Robinson's thesis showing test-set performance vs.\Nepoch count for some of the training runs. In most cases, performance\Npeaks at around 250 correct, after which performance decays to different\Ndegrees. The numbers given above are final performance figures after about\N3000 trials, not the peak performance obtained during the run.\N\N#### REFERENCES\N\N[Deterding89] D. H. Deterding, 1989, University of Cambridge, "Speaker\N Normalisation for Automatic Speech Recognition", submitted for PhD.\N\N[NiranjanFallside88] M. Niranjan and F. Fallside, 1988, Cambridge University\N Engineering Department, "Neural Networks and Radial Basis Functions in\N Classifying Static Speech Patterns", CUED/F-INFENG/TR.22.\N\N[RenalsRohwer89-ijcnn] Steve Renals and Richard Rohwer, "Phoneme\N Classification Experiments Using Radial Basis Functions", Submitted to\N the International Joint Conference on Neural Networks, Washington,\N 1989.\N\N[RabinerSchafer78] L. R. Rabiner and R. W. Schafer, Englewood Cliffs, New\N Jersey, 1978, Prentice Hall, "Digital Processing of Speech Signals".\N\N[PragerFallside88] R. W. Prager and F. Fallside, 1988, Cambridge University\N Engineering Department, "The Modified Kanerva Model for Automatic\N Speech Recognition", CUED/F-INFENG/TR.6.\N\N[BroomheadLowe88] D. Broomhead and D. Lowe, 1988, Royal Signals and Radar\N Establishment, Malvern, "Multi-variable Interpolation and Adaptive\N Networks", RSRE memo, #4148.\N\N[RobinsonNiranjanFallside88-tr] A. J. Robinson and M. Niranjan and F. \N Fallside, 1988, Cambridge University Engineering Department,\N "Generalising the Nodes of the Error Propagation Network",\N CUED/F-INFENG/TR.25.\N\N[Robinson89] A. J. Robinson, 1989, Cambridge University Engineering\N Department, "Dynamic Error Propagation Networks".\N\N[McCullochAinsworth88] N. McCulloch and W. A. Ainsworth, Proceedings of\N Speech'88, Edinburgh, 1988, "Speaker Independent Vowel Recognition\N using a Multi-Layer Perceptron".\N\N[RobinsonFallside88-neuro] A. J. Robinson and F. Fallside, 1988, Proceedings\N of nEuro'88, Paris, June, "A Dynamic Connectionist Model for Phoneme\N Recognition.\N\N\N#### Notes\N* This is version 2. Version 1 is hidden because it includes a feature dividing the data in train and test set. In OpenML this information is explicitly available in the corresponding task.