Abstract: Substantial pitch variations in artificial speech produced by dialing out a sequence of stored digital waveforms are made possible without significant distortion by varying pitch both by truncation or extension of pitch period waveforms, and by varying the dialout rate. In another aspect of the invention, pitch changes are made more natural by distributing each pitch change evenly over a large number of pitch periods during voiced phonemes.

Abstract: Artificial speech is produced with minimized memory requirements by using a bank of digital oscillators to produce voiced sounds by combining multiple harmonies of a fundamental-frequency sine wave, and using only one of these oscillators to reproduce stored nonsinusoidal waveforms for unvoiced sounds. Sufficient dynamic range is achieved with a minimum number of oscillators by generating only every other harmonic at the higher frequencies. All harmonics are derived from a single stored digitized sine wave by using stored sets of skip counts and amplitude codes corresponding to various voiced sounds to be produced.

Abstract: Improved speech sounds are produced from an all-software speech generating program on a computer which does not have a digital-to-analog converter by polling a first timer connected to the computer's speaker until a time-out is found, then applying a sample time value to a second timer so connected, and restarting the first timer. This results in the generation of a variable-duty-cycle carrier without interfering with the computer's operation. In an alternative embodiment, the first timer is used to generate an interrupt at regular intervals, and a sample time value from a storage buffer is then applied to the second timer.

Abstract: In a digital text-to-speech conversion system of the type usually contained in all-software form on a floppy disk, memory requirements for the storage of digitized waveform samples are reduced while speech quality is improved, by providing compression techniques and anti-distortion techniques which interact to provide clear speech at widely varying speeds with a minimum of memory. These techniques include using Huffman coding of first- or second-order differences, encoding only differences between successive waveforms where feasible, using a demi-diphone organization of the speech to allow use of the same instruction lists for several sounds, selectively deleting or repeating waveforms in the concatenation to vary speed without affecting pitch, and encoding waveforms linearly or anti-logarithmically for storage while converting the stored linear or anti-logarithmic codes to logarithmic codes such as .mu.-law codes upon retrieval.

Abstract: Clear speech can be produced on a DAC-less personal computer from an all-software speech-generating program with an unlimited vocabulary by applying to the computer's speaker a square-wave carrier signal of ultrasonic frequency, and varying the duty cycle of the carrier signal in accordance with digitized speech samples to produce an audio signal. The computations necessary to produce the speech samples are so timed by the program that the CPU is free to turn the speaker on and off between computations at precisely the right instant for the correct generation of a carrier having a constant ultrasonic frequency and a duty cycle determined by the value of the speech samples.

Abstract: A high-quality, real-time text-to-speech synthesizer system handles an unlimited vocabulary with a minimum of hardware by using a microcomputer-software-compatible time domain methodology which requires a minimum of memory and computational power. The system first compares test words to an exception dictionary. If the word is not found therein, the system applies standard pronunciation rules to the text word. In either instance, the text word is converted to a phoneme sequence. By the use of look-up tables addressed by pointers contained in a phoneme-and-transition matrix, the synthesizer translates the sequence of phonemes and transitions therebetween into sequences of small speech segments capable of being expressed in terms of repetitions of variable-length portions of short digitally stored waveforms.

Abstract: The memory needed for the digital storage of a waveform having an amplitude range of .+-.128 units is compacted by nearly 50% through the use of a 16-position table containing selected increments. Each waveform sample value is first predicted from at least two prior samples by linear prediction, and the predicted value is then incremented by the most nearly accurate increment in the table to obtain an approximation of the actual waveform value at the sample point. The appropriate increment for each sample can be defined by a 4-bit table address. The best-fit increment values for any given waveform can be calculated by a reiterative trial-and-error computation process. By the use of a best-fit table, inaccuracies in the waveform reproduction can be held below audible levels.