Time Interval Analyzer for Clipped Speech
Abstract Generated abstract
This paper describes an instrument for analyzing the distribution of time intervals between zero crossings in clipped speech, motivated by the use of such distributions in automatic speech recognition. The authors present a two channel analyzer that separately processes positive and negative rectangular pulses, measures their durations with a 22 kHz counting pulse generator, groups the resulting counts into 17 duration channels, and records frequency distributions with external counters. The design uses transistor module circuitry, diode transistor coincidence circuits, trigger counters, reset pulse shapers, and matrix decoders. The proposed setup provides one dimensional distributions for positive and negative clipped speech intervals and can support construction of a joint two dimensional distribution.
Full Text
ANALYZER OF TIME INTERVALS IN CLIPPED SPEECH
N. F. KUCHER, V. M. RAZIN
(Presented by the Department of Computer Engineering)
Since the publication of [1], many researchers have repeatedly turned to clipped speech. This is due to the fact that clipped speech, while having high intelligibility, possesses a substantially simpler waveform than natural speech and is a sequence of rectangular pulses of equal amplitude and varying duration.
According to Licklider’s hypothesis [1], information about the phonemic composition of speech is contained in the moments at which the speech signal crosses the zero axis. Clipped speech contains only this information. In this connection, it is of interest to find the distribution of the time intervals between zeros of clipped speech in order to study the possibility of using such distributions for automatic speech recognition.
This problem was first posed by Sakai and Inouye [2]. They constructed an instrument for analyzing the distributions of time intervals in clipped speech and obtained such distributions for Japanese phonemes. However, in many cases these distributions overlapped for different speakers. This indicates that the influence of the pronunciation of different speakers has a substantial effect on the process of automatic phoneme classification.
It is of interest to analyze clipped speech normalized by the method described in [3]. As a result of such normalization, the individual characteristics of the speaker—such as the fundamental frequency and the spectrum of the excitation function of the vocal-tract-forming system—are largely eliminated.
The authors developed an instrument for analyzing “artificial” clipped speech signals, the block diagram of which is presented in Fig. 1.
Speech clipped by the clipper described in [4] is fed to an input trigger, which is triggered by the leading and trailing edges of the clipped speech signal. As a result, two clipped speech signals shifted in phase by 180° are formed at the direct and inverse outputs of the trigger. The trigger outputs are connected to the corresponding channels of the analyzer, intended for separate analysis of the positive and negative pulses of clipped speech. As can be seen from Fig. 1, both channels are constructed according to completely identical circuits.
Let us consider the operation of one of the analyzer channels, for example, the channel of positive pulses. The pulse of the input trigger opens the time selector for the duration of the presence, at the trigger input, of a positive pulse of clipped speech. During this time, pulses from the counting-pulse generator pass through the time selector to the input of the seven-bit counter; their repetition frequency \(f_0\) is chosen in accordance with Kotelnikov’s theorem:
Fig. 1. Block diagram of a clipped-speech analyzer
\[ f_0 = 2\Delta F = 22\ \text{kHz}, \]
where
\[ \Delta F = 11\ \text{kHz} \]
is the spectral width of the clipped speech signal. Thus, the number counted by the counter proves to be proportional to the duration of the positive input pulse. During counting, the counter output, by means of the number-inhibit block, closed by a pulse from the inverse output of the trigger, is disconnected from the decoder, and there are no signals at the analyzer output.
At the moment when the positive pulse ends, the input trigger is reset; as a result, the inhibit block opens, and the binary seven-bit number from the counter output, proportional to the duration of the input pulse, is fed to the decoder input. The decoder performs two operations: conversion of the binary code into decimal and grouping of decimal numbers in accordance with the table. Such grouping coarsens the results of the analysis, but substantially reduces the time required to carry out the experiment.
Table
| Channel number | Number of pulses | Channel number | Number of pulses |
|---|---|---|---|
| 1 | 1 | 10 | 11–12 |
| 2 | 2 | 11 | 13–14 |
| 3 | 3 | 12 | 15–17 |
| 4 | 4 | 13 | 18–27 |
| 5 | 5 | 14 | 28–47 |
| 6 | 6 | 15 | 48–67 |
| 7 | 7 | 16 | 68–87 |
| 8 | 8 | 17 | 88–100 |
| 9 | 9–10 |
The presence of a pulse at a definite output of the decoder corresponds to the arrival, at the analyzer input, of a positive pulse of the corresponding duration.
The trailing edge of the pulse of the input trigger starts the reset-pulse shaping unit, which generates a reset pulse delayed by 20 μsec relative to the trailing edge of the trigger; this pulse brings the counter to the zero state. During this time, count pulses do not arrive at the counter, since the time selector is closed. Since 20 μsec elapse from the beginning of the opening of the inhibit unit until the counter is reset, the output pulse of the decoder also has a duration of 20 μsec.
When negative pulses arrive at the input of the analyzer, the second channel—the channel of negative pulses—comes into operation; its operation is completely analogous to that of the positive-pulse channel.
Figure 2 shows the functional diagram of the analyzer, which is at the same time also a circuit diagram, since any functional unit is a unified circuit element.
Fig. 2. Functional diagram of the analyzer
In designing the analyzer, a unified series of transistor flat modules of types P.12—P.13 was used.
The time selectors 3, 4 and the number-inhibit units 7, 8 (Fig. 2) are diode-transistor coincidence circuits. Counters 9 and 10 are ordinary trigger arrays. The reset-pulse shaping units 5 and 6, consisting of a delay device and a reset-pulse generator, are made in the form of a pair of waiting multivibrators connected in series, each of which generates a pulse of duration 20 μsec. The count-pulse generator 2 consists of an 88-kHz quartz generator and a self-oscillating multivibrator operating in frequency-division-by-4 mode. Decoders 11 and 12 are of the matrix type and are made with D220 silicon diodes.
The analysis of clipped speech is carried out in accordance with the block diagram of the experiment (Fig. 3) on an installation consisting of a tape recorder with an “endless” tape on which the speech sound under study is recorded,
analyzer and two counters connected to the outputs of both channels of the analyzer.
The distribution of time intervals is constructed according to the formula
\[ W(\tau_i)=\frac{1}{\Delta \tau_i}\,\frac{v_i}{\sum_{i=0}^{16} v_i}, \]
where \(i=0,1,2\ldots 16\),
\(\Delta \tau_i\) is the duration of the pulse at the analyzer input,
\(v_i\) is the number of pulses of duration \(\Delta \tau_i\).
Since the presence of a pulse at any output of the decoder corresponds to the appearance at the analyzer input of a pulse of a definite duration \(\Delta t\), the quantity \(\Delta t\) may be replaced by the number of the decoder output.
The quantity \(v_i\) is counted by means of a counter connected to the \(i\)-th output of the decoder.
One-dimensional distributions \(W_1(\tau_i)\) and \(W_2(\tau_i)\) are obtained independently for both channels of the analyzer; this, however, does not exclude the possibility of constructing the joint two-dimensional distribution of the time intervals of clipped speech by combining \(W_1(\tau_i)\) and \(W_2(\tau_i)\) in a two-dimensional coordinate system.
Fig. 3. Block diagram of the experiment
Figure labels: magnetic tape recorder; analyzer; counter 1; counter 2; outputs \(1\)–\(17\) and \(1'\)–\(17'\).
REFERENCES
-
Lisklider J. R., Pollack J. Effects of differentiation, integration and infinite peak clipping upon the intelligibility of Speech. JASA. v. 20, No. 1, 1948.
-
Sakai J., S. Inone. New Instruments and methods for speech analysis. JASA, v. 32, No. 4, 1960.
-
N. F. Kucher, V. M. Razin. On the question of the possibility of using “Artificial Larynx” apparatuses for speech research. Present collection.
-
N. F. Kucher, B. N. Epifantsev. Intonometer. Present collection.