Method of and device for the recognition, without previous training of connected words belonging to small vocabularies

Claim 2

Method of recognition of connected words belonging to small vocabularies providing for an initial step wherein a speech signal is digitized and is subdivided into intervals, in each interval an acoustic-phonetic analysis being made comprising : a linear-prediction encoding to calculate a normalized residual energy E R , formant frequencies Fr and an autocorrelation function of residual signal R x , wherefrom a first indication of voiced or unvoiced signal is extracted, depending on whether R x is greater or less than a threshold; a Fast Fourier Transform to calculate a total energy E T , a low-frequency energy E L , an intermediate frequency energy E M , and a high-frequency energy Eg wherefrom a value of spectral stability function SSF is calculated; providing also for an intermediate phase comprising the analysis of the trend of said intermediate frequency energy E M and of said formant frequencies Fr, so as to assign to interval sequences a division into nasal, semivowel, vowel classes with an indication of presence of dipthongs; providing also an end phase during which said word recognition is performed by analyzing a sequence of phonetic indications obtained during said intermediate phase by tree search algorithms of pattern matching of phonetic indications sequences, and dynamic programming, characterized in that said sequence of phonetic indications consists only of a subdivision of the total sequence of intervals into phonetic classes comprising said nasal, semivowel, vowel classes with diphtong presence, and silence, voiced fricative, unvoiced fricative, plosive, affricate classes; said silence class being detected during intervals in which said normalized residual energy E R exceeds a value of reference energy E RIF calculated on the basis of said total energy E T ; the beginning of one of said fricative, plosive, affricate classes being detected in the interval in which said spectral stability function exceeds a spectral stability threshold, their identification being executed by analyzing, in a number of intervals (-N, +N) preceding and following the interval of said beginning, the presence of silence class, the comparison of said total energy E T with energy thresholds (E 1 , E 2 ), and a subdivision into four voicing degrees of said intervals.

Claim 3

Method according to claim 1, characterized in that said value of reference energy E RIF is given by the following formula: where a is a constant, E T is said total energy, E TI is a mean total energy extending over a number of initial intervals.

Claim 4

Method according to claim 1 or 2, characterized in that a ratio R between low frequency and high frequency energy E L , Eg is calculated and then compared with a threshold, and said voicing degrees are: - voiced, if said first indication is of a voiced sound and said ratio exceeds the threshold; - quasi-voiced, if the first indication is of unvoiced sound and said ratio exceeds the threshold; - quasi-unvoiced, if the first indication is of voiced sound and said ratio does not exceed the threshold; - unvoiced, if the first indication is of unvoiced sound and said ratio does not exceed the threshold.

Claim 5

Method according to claim 3, characterized in that a voiced fricative class is identified if most of said preceding intervals have been classified as silence, said total energy is less than a first of said energy thresholds (E 2 ) in the interval of beginning, and most of the subsequent intervals have been assigned either a voiced or quasi-voiced degree, or quasi-voiced or quasi-unvoiced degree; or even if most of the preceding intervals have been assigned a voiced or quasi-voiced degree, and most of the subsequent ones either a voiced or quasi-voiced degree, or quasi-voiced or quasi-unvoiced degree; said voiced fricative class lasting up to the interval in which total energy E T exceeds said first threshold.

Claim 6

Method according to claim 3 or 4, characterized in that an unvoiced fricative class is identified if most of the subsequent intervals have been assigned an unvoiced or quasi-unvoiced degree, and if most of the preceding intervals have been assigned a voiced or quasi-voiced degree or the silence class, and in most of the subsequent intervals the total energy E T is less than a second of said energy thresholds (E l ), said unvoiced fricative class lasting up to the interval which is assigned the voiced or quasi voiced degree.

Claim 7

Method according to any of claims 3 to 5, characterized in that a plosive class is identified if most of the preceding intervals have been assigned a silence class, during the interval of beginning the total energy is greater than the first energy threshold (E 2 ), and if most of the subsequent intervals have been assigned voiced or quasi-voiced degree, said plosive class identifying the interval of beginning alone, or, if quasi-voiced or quasi-unvoiced degree has been allotted, said subsequent intervals.

Claim 8

Method according to any of claims 3 to 6, characterized in that an affricate class is identified if most of the preceding intervals have been assigned the silence class, the subsequent intervals unvoiced or quasi-unvoiced degree, and in most of the subsequent intervals total energy is greater than said second energy threshold (E l ), said affricate class lasting until the interval which is reallot ted the voiced or quasi-voiced degree.

Claim 9

Device for implementing the method of any claim from 1 to 7, comprising an analog-to-digital converter (AD) of the speech signal subdivided into intervals, a circuit (LPC) for computing linear prediction coefficients (ai, pi) relevant to digital samples of each interval, followed by circuits for computing said normalized residual energy E R , said formant frequencies Fr, said autocorrelation function of the residual signal R x and said first indication of voiced or unvoiced sound, a circuit (FFT) for calculating the Fast Fourier Transform of digital samples of each interval, followed by a circuit (CEN) for computing said total, low-frequency, intermediate-frequency and high-frequency energies (E T , E L , E M , Eg), a circuit (SSF) for computing said spectral stability function, emitting a signal (13) indicating whether said spectral stability threshold has been exceeded, a circuit (SMOT) for determining the values of a smoothed intermediate-frequency energy function E' M , and of its maxima and minima, and a processing circuit (ELB) for said end phase, characterized in that it further comprises: - a circuit (SIL) for detecting said silence class, which receives the values of residual energy E R and total energy E T ; computes said reference energy E RIF and compares it with said residual energy E R ; - a circuit (VNV) determining said subdivision into four voicing degrees, on the basis of said first indication of voiced or unvoiced sound, and of the computation of said ratio between the energies at low and high frequencies E L , E H ; - a circuit (COMP) for the comparison of said total energy ET with said first and second energy thresholds (E 1 , E 2 ); - a first register (RG1) for combining data words, one per each interval, composed of said silence class, said formant frequencies Fr, said voicing degrees, said signal outgoing from said circuit (COMP) for the total energy comparison, the signal indicating whether the spectral energy threhsold has been exceeded, the values of the smoothed intermediate-frequency energy function E' M and maximum and minimum indications; - a memory (MEMI) for the temporary storage of the last M data. words; - a circuit (CLSS) for determining said phonetic classes upon the analysis of said data words it receives via a data bus (19) form said memory (MEM1), which it supplies with the addresses on an address bus (20), said classes being supplied to said processing circuit (ELB) for the end phase, with a signal for the interval identification.

Claim 10

Device as in claim 8, characterized in that said circuit (CLSS) for determining the phonetic classes essentially comprises: - a first unit (IND1) for the sequential addressing of said memory (MEM1) during the reading, said first unit being stop ped by a voice signal (Q), and restarted from the value present on the address bus (20) of said memory by a first recovery signal, and supplying, when active, a sequence of pulses of interval identification; - a second unit (IND2) for the sequential adressing of said memory (MEMI) during reading, said second unit carrying out at each activation a first and a second addressing and disabling a third unit (IND3), the first addressing beginning from N positions before the address present at the activation on said address bus (20) and addressing 2N+1 subsequent positions, the second addressing starting from said address present at the activation on the address bus (20) and ending when said second unit receives a second inhibiting signal (32); - said third unit (IND3) addressing during reading said memory (MEMI) starting, at each activation, from the address present on said address bus (20), disabling said second unit (IND2) and being stopped by a third inhibiting signal (35); - a second register (RQ), which temporarily stores and supplies to the output a field of said data words, read in said memory (MEMI), carrying the silence class (Q) or said voice signal (Q), said silence class being supplied to said processing circuit (ELB) for the end phase; - a third register (RSF) which, when activated by said voice signal (Q), temporarily stores a field of said data words, read in the memory (MEM1), carrying said signal indicating whether said spectral stability threshold has been exceeded, and supplies to the output the activation signal (SF, SF), for said second (IND2) or said third (IND3) unit depending on whether said field indicates threshold exceeded or not exceeded threshold; - a circuit (VNS) for determining nasal, semivowel, vowel classes and the presence of diphtongs on the basis of the analysis of the fields of said data words present on the data bus (19) carrying said formant frequencies Fr and the values of the smoothed intermediate-frequency energy function E' M and the indications of maxima and minima, said circuit (VNS) being activated when said third unit (IND3) is activated and supplying it with the control signals relevant to the memory (MEM1) addressing, also supplying a sequence of pulses identifying intervals, and being disabled when in a data word there is present again said silence class or the indication of spectral stability threshold exceeded, thus supplying said third inhibiting signal (35); said nasal, semivowel, vowel classes and diphtong presence being supplied to said processing circuit (ELB) for the end phase; - a first logic circuit (FPA) which generates condition signals (37) for the identification of fricative, plosive, affricate classes, on the basis of the analysis of the fields of said data words present on the data bus (19), carrying the silence class, voicing degrees, the comparisons of the total energy with said first and second threshold (E 2 , E l ); said first logic circuit being activated by said first addressing supplied by the second unit (IND2); - a combinatory logic (LGC) of said condition signals (37), emitting an identification signal (38) of voiced-fricative or unvoiced fricative, or plosive or affricate classes; - a second logic circuit (DUR) for determining the number of intervals labelled by fricative, plosive, affricate classes, which receives said identification signal (38) from said combinatory logic, the fields of said data words carrying voicing degrees and the comparisons of the total energy with the thresholds, which is activated by said second addressing supplied by said second unit (IND2), which supplies one of said fricative, plosive, affricate classes to said processing circuit (ELB) for the end phase, and also supplies a sequence of pulses identifying the intervals, and, at the end of the sequence, said second inhibiting signal (32) and said third inhibiting signal (35) being also said first recovery signal; - a logic gate (P6) supplying said signal (40) for interval identification to the processing circuit (ELB) for the end phase on the basis of the sequences of pulses identifying the intervals it receives from the first unit (INDl), from the circuit (VNS) for the determination of nasal, semivowel and vowel classes, and from said second logic circuit (DUR).

Claim 11

Device as in claim 9, characterized in that said first ,logic circuit (FPA) basically comprises: - a first counter (CA) of the number of preceding intervals in which the silence class is present; - a second counter (CB) of the number of preceding intervals having a voiced or quasi-voiced degree; - a third counter (CC) of the number of subsequent intervals having an unvoiced or quasi-unvoiced degree; - a fourth counter (CD) of the number of subsequent intervals having a voiced or quasi-voiced degree; - a fifth counter (CE) of the number of subsequent intervals having a quasi-voiced or quasi-unvoiced degree; - a sixth and a seventh counters (CF, CG) of the number of subsequent intervals in which total energy E T does not exceed or exceeds respectively said second threshold (E l ); said first to seventh counters supplying an active logic level when exceeding half the maximum counting value; - an eigth and ninth counters (CH, CI) which supply an active logic level if in said initial interval the total energy E T is respectively less or greater than said first threshold (E2); - a circuit (DIN) allotting the pulses of said first addressing to said counters as synchronism signals; - a fourth register (RG3) which receives the outputs of said counters and supplies them on the output.as condition signals (37) at the end of said first addressing.

Claim 12

Device as in claim 10, characterized in that said combinatory logic (LGC) emits: - a signal of voiced-fricative class (Fvl) if the outputs are active of the following counters: first and fourth and eight, or first and fifth and eight, or second and fourth or yet second and fifth; - a signal of unvoiced-fricative class (Fnl) if the outputs are active of the following counters: first and third and sixth, or second and third; - a first signal of plosive class (P 1 ) if the outputs of said first, fourth and ninth counters are active; - a second signal of plosive class (P 2 ) if the outputs of said first, fifth and eight counters are active; - a signal of affricate class (A 1 ) if the outputs of said first, third and seventh counters are active.

Claim 13

Device as in claim 11, characterized in that said second logic circuit (DUR) essentially comprises: - a fifth register (RG4) which emits said voiced-fricative class (F v ) if it receives said signal of voiced-fricative class (F vl ) and as long as total energy ET is less than said first threshold (E 2 ); - a sixth register (RGS) which emits said unvoiced-fricative class (F n ) if it receives said signal of unvoiced-fricative class (F nl ) and as long as the unvoiced or quasi-unvoiced degree is present; - a seventh register (RG6) which emits said affricate class (A) if it received said affricate class signal (A 1 ) and as long as unvoiced or quasi-unvoiced degree is present; - an eighth register (RG7) which emits an active logic level if it receives said second signal of plosive class (P 2 ) and as long as the quasi-voiced or quasi-unvoiced degree is present; said fifth, sixth, seventh, eighth registers being synchronized by the pulses of said second addressing; - a logic gate (P15) which emits said plosive class (P) if the output of said eighth register (RG7) is active or if it receives said first signal of plosive class (P 1 ); - a ninth register (RG9) which supplies the pulses of said second addressing (39) as a sequence indentifying the intervals if and as long as one of said fricative, plosive, affricate classes is present; - a tenth register (RG8) which emits said second inhibiting signal (32) when said fifth, sixth, seventh, eighth registers are disabled.