Search

A fast nonlinear time alignment method is presented, which is based on a preprocessing of the normalized speech spectrogram by means of a segmentation of the trace in the spectral feature space. After such trace segmentation the patterns have a fixed format and allow for a subsequent classification with a distance measure which is obtained from conventional dynamic programming with extreme constraints. Since, due to the trace segmentation preprocessing, these extreme constraints can be applied without performance degradation, the described method offers savings in computing time by a factor of 10 or more as compared to conventional dynamic programming. As a side benefit, reference pattern memory savings by a factor of 3 or more are obtained.

Some results obtained when the recognition vocabulary size of a phoneme-based speaker-dependent continuous-speech recognizer was increased from 1000 to 10000 words are reported. The potential search space increased from 46000 to 516000 states without problems for the data-driven search. Increasing the recognition vocabulary by a factor of 10 (from a perplexity of 917 to 9686) increased the word error rate by a factor of two (from 21.8% to 43.1%). Phoneme models were tested with both discrete probabilities and continuous mixture densities. The mixture density models performed better; moreover, they saved about half of the search costs. A language model was found to be very important for a larger vocabulary size. With a test set perplexity of 388 (i.e. a reduction by a factor of 25 compared to the case without a bigram model) the error rate decreased by a factor of 2.4. In order to check how meaningful perplexity is for the prediction of the system's performance, a stochastic language model was constructed with a perplexity of 1000, the size of the vocabulary used in previous experiments, and about the same error rate was obtained. >

We present a particular way of connecting syntax, semantics, and speech recognition, which uses a Single syntactic-semantic representation both as a backbone for the speech recognition and semantic interpretation modules. A spoken-language database query system is described which was developed in this framework with a commercial speech recognition development toolkit.

Durch eine mehrstufige Klassifikation der im Rahmen des AUROS-Sprecher-erkennungssystems aufgenommenen Sprachproben zunachst mit einem linearen, dann mit einem stuckweise linearen Abstands-Klassifikator (Minimum Distance, Nearest Neighbour), konnte eine Verbesserung der Verifikationsergebnisse bei gleichzeitiger deutlicher Verringerung von Rechenzeit und Speicherplatzbedarf je Sprecher erzielt werden.

In unserem Labor wurde in den vergangenen zwei Jahren im Rahmen des Verbundprojektes ‚SPICOS‘ ein sprecherabhangiges System zur Erkennung von gesprochenen Daten­bankanfragen entwickelt. Einzelheiten der Algorithmen und der verwendeten Wissens­quellen wurden bereits an anderer Stelle beschrieben ([1],[2],[3],[4]). In diesem Beitrag wird versucht, die Ergebnisse von einigen Erkennungsexperimenten auszuwer­ten. Der Aufbau dieses Artikels ist wie folgt: im Abschnitt 2 werden die wesent­lichen Eigenschaften des Suchverfahrens und die Bedingungen fur 3 Arten von Experi­menten beschrieben, die in den darauf folgenden Abschnitten im Einzelnen behandelt werden.

PHD. 80-037 ABSTRACT In a device for adjusting a movable or electro-acoustic sound transducer an optical duct is connected the sound transducer, which only enables an object, such as for example a display table or a data display unit to be viewed completely in a specific position. This position is unambiguously reproducible and thus also the position of the sound transducer or microphone relative to the mouth of the speaker. The optical duct may be constituted by diaphragms or a tube, as the case may be with intermediate walls, or a phase grating or amplitude grating. In a special case, the object is a display device and the device in accordance with the invention further comprises a device for generating variable data on a display. Since during the speaker identification process the device for the gene-ration of variable data projects consecutive instructions on the display device, the speech recognition process can thus be controlled automatically. The speaker to be identified is then constantly forced to keep his mouth in the correct position relative to the microphone.

An improved method of verifying the speaker from whom long term spectral characteristics have been derived during a learning phase. Mean values and distance thresholds have been calculated and stored. Improved recognition is obtained by incorporating into the stored values additional values of speech samples in ranges situated outside the distance threshold value. When, during verification, a speech sample not situated within the distance threshold value is encountered its distance from the additionally stored speech samples is successively determined and compared with the separate distance threshold value. The speaker is rejected only if the speech sample to be verified is not situated at a sufficiently small distance from any of the additionally stored speech samples. The method thus increases the recognition range to better approximate the actual distribution of speech samples of the learning phase.