Abstract: A speech encoder uses a soft interpolation decision for spectral parameters. For each frame, the encoder first calculates the residual energy for interpolated spectral parameters, and then calculates the residual energy for non-interpolated spectral parameters. The encoder then compares these residual energy calculations. If the encoder determines that the interpolated spectral parameters yields the lowest residual energy, it indicates to a far-end decoder to use the interpolated values for the current frame. Otherwise, it indicates to the far-end decoder to use the non-interpolated values for the current frame. The encoder signals the far-end decoder as to which spectral parameters (interpolated or non-interpolated values) to use by encoding and transmitting a special signalling bit.

Abstract: Lag information for use in a speech coder is developed by estimating lag values for the various subframes (201) of a speech coding frame (200) of information, and by then selecting lag values for each subframe that are both closely corresponding to the estimated lag values and that also observe the restrictions of a selected delta-coding routine. When a plurality of candidate sets of such information have been developed, they are compared against one another to identify that set which appears to provide the best set of lag values. This information is then available for framing and transmission. In one embodiment, the sets of candidate values are also selected to ensure provision for subsequent adjustment in either a positive or negative direction.

Abstract: An adaptive spectral postfilter in a synthesized speech platform has a denominator characteristic that corresponds to a preceding LPC filter stage, and a numerator characteristic that is developed as a function of the denominator characteristic through application of spectral smoothing techniques. This allows the numerator to track the denominator without the introduction of spectral distortion that would otherwise affect the processing in an adverse way.

Abstract: In a Viterbi Algorithm decoder (204) as used to decode convolutionally encoded information, reliability information is developed for various path discard decisions made within the Viterbi Algorithm. These decisions are made for discard opportunities that impact one or more error detection windows (601). Based upon these metrics, a reliability factor sequence can be provided and compared against a fixed (or varying) threshold. When unreliability appears, appropriate action can be taken. For example, all of the information can be discarded, or only certain portions of the information can be discarded, as appropriate to the particular application.

Abstract: A channel bank speech synthesizer for reconstructing speech from externally-generated acoustic feature information without using externally-generated voicing or pitch information is disclosed. An N-channel pitch-excited channel bank synthesizer (340) is provided having a first low-frequency group of channel gain values (1 to M) and a second high-frequency group of channel gain values (+1 to N). The first group controls a first group of amplitude modulators (950) excited by a periodic pitch pulse source (920), and the second group controls amplitude modulators excited by a noise source (930). Both groups of modulated excitation signals are applied to the bandpass filters (960) to reconstruct the speech channels, and then combined at the summation network (970) to form a reconstructed synthesized speech signal. Additionally, the pitch pulse source (920) varies the pitch pulse period such that the pitch pulse rate decreases over the length of the word.

Abstract: A continuous speech recognition system employs a grammar tree of alternative potentially recognized word paths. A technique of tracing back through the grammar tree is utilized in determining which partial word path is common to all potential word paths. The common partial word path is deleted and words corresponding to the deleted partial word path are output as recognized words.

Abstract: Word spotting in a speech recognition system without predetermining the endpoints of the input speech. The invention is intended to be implemented in a system which has word templates stored in template memory, with the system being capable of accumulating distance measures for states within each word template. The following steps are used to generate a measure of similarity between a subset of the input frames and a word template. The steps are: a) recording a beginning input frame number for each state to identify the potential beginning of the word; b) accumulating distance measures for at least one state for each input frame; c) normalizing the distance measures by substracting a normalization amount from each distance measure; d) recording normalization information corresponding to the normalization amount for each input frame; and e) determining a similarity measure between the word template and a subset of input frames after a given input frame has been processed.

Abstract: A method and arrangement for a speech recognition system employs channel bank information to represent speech. The method considers background noise included with the speech. The method includes determining three energy levels for each channel the first representative of background noise energy, the second representative of the input frame energy and the third representative of the word template frame energy. Values representing energy level differentials are assigned at each channel. If the second energy level is less than the first energy level, then a predetermined constant value is assigned at the particular channel. These values are combined to generate a distance measure depicting the similarity between the two frames.

Abstract: The present invention describes a method and arrangement for reducing a sequence of initial frames into a reduced set of representative frames by combining the initial frames into a plurality of representative frames, the combining process including generating a distortion measure associated with each representative frame and comparing each distortion measure to a distortion threshold. From these representative frames, a set of mutually exclusive frames is determined to minimize the number of representative frames, whereby each representative frame in the set represents a unique set of contiguous initial frames and has an associated distortion measure which does not exceed the distortion threshold.

Abstract: An improved excitation vector generation and search technique (FIG. 1) is described for a code-excited linear prediction (CELP) speech coder (100) using a codebook memory of excitation code vectors. A set of M basis vectors v.sub.m (n) are used along with the excitation signal codewords (i) to generate the codebook of excitation vectors u.sub.i (n) according to a "vector sum" technique (120) of converting stored selector codewords into a plurality of interim data signals, multiplying the set of M basis vectors by the interim data signals, and summing the resultant vectors to produce the set of 2.sup.M codebook vectors. Only M basis vectors need to be stored in memory (114), as opposed to all 2.sup.M code vectors.

Abstract: A user-interactive speech recognition control system is disclosed for recognizing a complete sequence of keywords (e.g., a telephone number such as 123-4567) via entering, verifying, and editing variable-length utterance strings (e.g., 1-2-3; 4-5; 6-7) separated by the user-defined placement of pauses. The device controller (120) utilizes timers (124) to monitor the pause time between partial-sequence digit strings recognized by the speech recognizer (110). When a string of digits is followed by a predetermined pause time interval, the recognized digits will be replied via the speech synthesizer (130). An additional string of digits can then be entered, and only the subsequent string will be replied after the next pause. Furthermore, the user has the flexibility to correct only the last digit string entered, or the entire sequence. Hence, if there is an error in only one digit, the erroneous digit string can be corrected without having to re-enter the entire digit sequence.

Abstract: An improved excitation vector generation and search technique (FIG. 1) is described for a code-excited linear prediction (CELP) speech coder (100) using a codebook of excitation code vectors. A set of M basis vectors V.sub.m (n) are used along with the excitation signal codewords (i) to generate the codebook of excitation vectors u.sub.i (n) according to a "vector sum" technique (120) of converting the selector codewords into a plurality of interim data signals, multiplying the set of M basis vectors by the interim data signals, and summing the resultant vectors to produce the set of 2.sup.M codebook vectors. The entire codebook of 2.sup.M possible excitation vectors is efficiently searched by using the vector sum generation technique with the M basis vectors--without ever having to generate and evaluate each of the 2.sup.M code vectors themselves. Furthermore, only M basis vectors need to be stored in memory (114), as opposed to all 2.sup.M code vectors.

Abstract: An improved noise suppression system (800) is disclosed which performs speech quality enhancement upon the speech-plus-noise signal available at the input (205) to generate a clean speech signal at the output (265) by spectral gain modification. The improvements of the present invention include the addition of a signal-to-noise ratio (SNR) threshold mechanism (830) to reduce background noise flutter by offsetting the gain rise of the gain tables until a certain SNR threshold is reached, the use of a voice metric calculator (810) to produce more accurate background noise estimates via performing the update decision based on the overall voice-like characteristics in the channels and the time interval since the last update, and the use of a channel SNR modifier (820) to provide immunity to narrowband noise bursts through modification of the SNR estimates based on the voice metric calculation and the channel energies.

Abstract: Described herein, is an arrangement and method for processing speech information in a speech recognition system (300). In such a system where the speech information is depicted as words, each word representing a sequence of frames (510) and where the recognition system has means (120) for comparing present input speech to a word template, the word template stored in template memory and derived from one or more previous input word, the present invention is best employed. The invention describes combining contiguous acoustically similar frames (512) derived from the previous input word or words into representative frames to form a corresponding reduced word template, storing the reduced word template in template memory in an efficient manner, and comparing frames of the present input speech to the representative frames of the reduced word template according to the number of frames combined in the representative frames of the reduced word template.

Abstract: Disclosed is a method for generating word templates for a speech recognition system. It is used where speech is represented by data in frames of equal time intervals. The method includes generating an interim template, generating a time alignment path between the interim template and a token, mapping frames from the interim template and the token along the time alignment path onto an averaged time axis, and combining data associated with the mapped frames to produce composite frames representative of the final word template. The method realizes advantages of reduced memory usage and a realistic data average from each contributing averaged word.

Abstract: An improved hands-free user-interactive control and dialing system is disclosed for use with a speech communications device. The control system (400) includes a dynamic noise suppressor (410), a speech recognizer (420) for implementing voice-control, a device controller (430) responsive to the speech recognizer for controlling operating parameters of the speech communications device (450) and for producing status information representing the operating status of the device, and a speech synthesizer (440) for providing reply information to the user as to the speech communications device operating status. In a mobile radiotelephone application, the spectral subtraction noise suppressor (414) is configured to improve the performance of the speech recognizer (424), the voice quality of the transmitted audio (417), and the audio switching operation of the vehicular speakerphone (460).

Abstract: An improved background noise estimator (320) is disclosed for use with a noise suppression system (300) for generating an estimate of the background noise power spectral density provided to noise suppressor (310), which performs speech quality enhancement upon the pre-processed speech-plus-noise signal available at the input to generate a clean post-processed speech signal at the output. Background noise estimator (320) utilizes an energy valley detector based upon post-processed speech to perform the speech/noise classification, and a noise spectral estimator based upon pre-processed speech to generate an estimate of the background noise power spectral density. As a result, the background noise estimate supplied to the noise suppressor is a more accurate measurement of the background noise energy, since it is performed during a more accurate determination of the occurrences of pauses in the speech.

Abstract: An automatic gain selector is disclosed for use with a noise suppression system which performs speech quality enhancement upon a noisy speech signal available at the input to generate a noise-suppressed speech signal at the output by spectral gain modification. The channel gain controller (240) of the present invention produces a modification signal (245), comprised of individual channel gain values, for application to a channel gain modifier (250). A particular gain table set is automatically selected from one of a plurality of gain tables (450) by a selector switch (470) and a noise level quantizer (440) in response to a multi-channel noise parameter, such as the overall average background noise level of the input signal. Then the individual channel gain values (455) are obtained from the particular gain table set in response to the individual channel signal-to-noise ratio estimate (235).

Abstract: An improved noise suppression system (400) is disclosed which performs speech quality enhancement upon speech-plus-noise signal available at the input (205) to generate a clean speech signal at the output (265) by spectral gain modification. The noise suppression system of the present invention includes a background noise estimator (420) which generates and stores an estimate of the background noise power spectral density based upon pre-processed speech (215), as determined by the detected minima of the post-processed speech energy level. This post-processed speech (255) may be obtained directly from the output of the noise suppression system, or may be simulated by multiplying the pre-processed speech energy (225) by the channel gain values of the modification signal (245). This technique of implementing post-processed signal to generate the background noise estimate (325) provides a more accurate measurement of the background noise energy since it is based upon much cleaner speech signal.

Abstract: An improved method and means of determining reflection coefficients that characterize an electrical signal that obtains characteristics of an all-zero inverse lattice filter. The reflection coefficients are obtained by filtering the signal, sample the filtered signal, obtaining the elements of a correlation array from the samples, initializing values of arrays forward residuals, backward residuals, and cross correlation of residuals, combining array elements to obtain a first reflection coefficient, removing from the forward, backward and cross-correlation arrays the effect of the first reflection coefficient, calculating from the revised arrays a second coefficient, and repeating the calculations to the desired order. In a second embodiment of the present invention, samples are selected from the digitized signal and multiplied by a windowing function.