Voiced Portion Research Articles

Separation of Singing Voice Using Nonnegative Matrix Partial Co-Factorization for Singer Identification

In order to improve the performance of singer identification, we propose a system to separate singing voice from music accompaniment for monaural recordings. Our system consists of two key stages. The first stage exploits the nonnegative matrix partial co-factorization (NMPCF), which is a joint matrix decomposition integrating prior knowledge of singing voice and pure accompaniment to separate the mixture signal into singing voice portion and accompaniment portion. In the second stage, based on the separated singing voice obtained by the first stage, the pitches of singing voice are first estimated and then the harmonic components of singing voice can be distinguished. For a frame, the distinguished harmonic components are regarded as reliable while other frequency components unreliable, thus the spectrum is incomplete. With those harmonic components, the complete spectrums of singing voice can be reconstructed by a missing feature method, spectrum reconstruction, obtaining a refined signal with more clean singing voice. Experimental results demonstrate that, from the point view of source separation, the singing voice refinement can further improve ΔSNR in contrast with the singing voice separation using NMPCF, while for the point view of singer identification, the singing voice separated by NMPCF is more appropriate than the refined singing voice.

Pick‐and‐place application development using voice and visual commands

PurposeThe purpose of this paper is to design an interactive industrial robotic system which can be used to assist a “layperson” in re‐casting a generic pick‐and‐place application. A user can program a pick‐and‐place application simply by pointing to objects in the work area and speaking simple and intuitive natural language commands.Design/methodology/approachThe system was implemented in C# using the EMGU wrapper classes for OpenCV as well as the MS Speech Recognition API. The target language to be recognized was modelled using traditional augmented transition networks which were implemented as XML Grammars. The authors developed an original finger‐pointing algorithm using a unique combination of standard morphological and image processing techniques. Recognized voice commands trigger the vision component to capture what a user is pointing at. If the specified action requires robot movement, the required information is sent to the robot control component of the system, which then transmits the commands to the robot controller for execution.FindingsThe voice portion of the system was tested on the factory floor in a “typical” manufacturing environment, which was right at the maximum allowable average decibel level specified by OSHA. The findings show that a modern/standard MS Speech API voice recognition system can achieve a 100 per cent accuracy of simple commands; although at the noisy levels of 89 decibels on average, every one out of six commands had to be repeated. The vision component was test of 72 test subjects who had no prior knowledge of this work. The system accurately recognized what the test subjects were pointing at 95 per cent of the time within five seconds of hand readjusting.Research limitations/implicationsThe vision component suffers from the “typical” problems: very shiny surfaces can cause problems; very poor contrast between the pointing hand and the background; and occlusions. Currently the system can only handle a limited amount of depth recovery using a spring mounted gripper. A second camera (future work) needs to be incorporated in order to handle large depth variations in the work area.Practical implicationsThis system could have a huge impact on how factory floor workers interact with robotic equipment.Originality/valueThe testing of the voice system on a factory floor, although simple, is very important. It proves the viability of this component of the system and debunks arguments that factories are simply too noisy for current voice technology. The unique finger‐pointing algorithm developed by the authors is also an important contribution to the field. In particular, the manner in which the pointing vector was constructed. Furthermore, very few papers report results of non‐experts using their pointing algorithms. The paper reports concrete results that show the system is intuitive and user friendly to “laypersons”.

The effect of segmental structure on the production of Japanese pitch‐accent

Japanese pitch‐accent is a fall in fundamental frequency (F0) after the accented mora. In current models of Japanese pitch‐accent contrast, accented and unaccented pitch patterns are predicted to be similar across syllable structure type. This study provides evidence that the segmental make‐up of words affects accentual patterns in Tokyo Japanese in terms of their F0 contours. Six native speakers produced accented and unaccented words that varied in segmental type of the word‐medial mora (N=864). It was found that the pitch‐accent contrast was produced differently in words with a short vowel in the first syllable, CVCVCV and CVC:V words (V‐group), on one hand, and words with a longer voiced portion of the first syllable, CV:CV and CVNCV words (V:/VN‐group), on the other hand, p>0.01. In the accented condition, the average F0 fall was 50 Hz for all, but the F0 movement within the first syllable was different between the two word groups. In the unaccented condition, word‐level, as well as intrasyllabic differences, were observed in F0 contours. F0 was rising in the V‐group, whereas it was flat in the V:/VN‐group. These results indicated that pitch patterns varied by segmental structure, more so for unaccented words.

Voice activated text strings for electronic devices

A system and method for associating text with a voice command and entering the associated text into an electronic device is disclosed. In one embodiment, the invention is a system for entering text into an electronic device, comprising, an input element for entering a text portion into the electronic device, a first memory for storing the text portion, a software code segment for recording a voice portion into the electronic device, and a second memory for storing the voice portion. The system also includes a software code segment for associating the voice portion with the text portion, a software code segment for associating the text portion with information to be entered in the electronic device, and voice recognition software to recall the text portion that is associated with the voice portion, where the input element inserts the text portion associated with the information to be entered into the electronic device.

Speech recognition with text generation from portions of voice data preselected by manual-input commands

A computer reads a voice speech recognition program from a first recording medium, and reads voice data from a second recording medium, and causes a CPU in the computer to recognize speech represented by the read voice data according to the speech recognition program, convert the result of speech recognition into text data, and display the converted text data on a display unit. A check mark button used by a speaker designates a portion of voice data, which is input through a microphone, corresponding to an unnecessary word or the like. The portion of the voice data in which a check mark is inscribed is not regarded as an object of speech recognition. Only the other portion of voice data in which the check mark is not inscribed is regarded as an object of speech recognition, and speech recognition is thus carried out. Furthermore, the sound level of a voice portion of voice data is rated. The gain of the voice data is adjusted according to the rated level. On the basis of the voice data whose sound level has been adjusted, speech recognition is carried out.

Full duplex speakerphone

First Page

Two types of voiceless nasals

Many languages in South East Asia have voiceless nasal consonants that contrast with their voiced counterparts. What has not been reported previously is that there are two distinct types of voiceless nasals. We will begin by considering the more well-known type, found in languages such as Burmese. These voiceless nasals are usually said to have an open glottis for most of the articulation, but some voicing for the period just before the stricture is broken (Ladefoged 1971: 11). In this view they have two parts. The first is necessary to distinguish them from their voiced counterparts. The second distinguishes one voiceless nasal from another, by making the place of articulation more apparent; the voiced offglide from the nasal into the vowel has formant transitions that are characteristic of each place of articulation. Dantsuji has shown that, in addition, the voiced portion of the voiceless nasals in Burmese “includes significant cues which can distinguish points of articulation” (Dantsuji 1986: 10). Despite the small voiced portion, in most phonological treatments of these sounds, a voiceless nasal consonant is considered not as a sequence of two items, a voiceless segment followed by a voiced segment, but as a single voiceless unit with a low level phonetic rule inserting the voicing towards the end. Burmese is a textbook example, with forms as shown in Table 1 (from Maddieson 1984; also in Ladefoged 1993: 282).

Explorations in speaker normalization

Some preliminary efforts are reported at speaker normalization based on the assumption that listeners are tacitly aware (a) that they are listening to an instrument (the human vocal tract) which when uniform (unconstricted) should have resonances whose lowest frequencies are spaced according to the ratios 1:3:5:7, etc., (b) that the exact frequencies of these resonances depend on the length of the vocal tract, and (c) that what is relevant in resonances from a nonuniform tract is the ratio of these lowest resonances to their uniform value. There are probably several cues the listener could use to estimate the size of the speaker; for starters, we use the mean (long‐term) F4. The reference resonances for F1r, F2r, F3r, i.e., those from the “uniform” tract, are 1/7, 3/7, and 5/7, respectively, of mean F4. The normalization consists in converting measured formant frequencies, F1m, etc., into the ratios of log F1m/log F1r. The results of applying this normalization to the entire voiced portion of vowels in a ...

Structure and duration of vowels together specify fricative voicing.

The durations of vowels specify several types of linguistic information. Thus multiple durational cues may be decoded from the same temporal interval of the signal, a situation which should create perceptual ambiguity. Four perceptual studies of postvocalic fricative voicing examined how information in the duration and dynamic structure of the vowel might combine to resolve these ambiguities. Stimuli were made by varying the duration and structure of vowels from utterances of /jus/ and /juz/. In the first experiment vowel structure differences influenced voicing judgments by 30%–40%. The second experiment showed that the offset characteristics of the /juz/ vowel were not the major vowel structure cues in the first experiment. The third experiment examined identification and discrimination of stimuli with neutral vocalic voicing cues which differed only in friction duration. Even with these stimuli, friction duration was a weak and poorly discriminated voicing cue. The fourth experiment showed that variations in the structure and duration of vowels are specific to the voicing contrast and minimally influence vowel identification. Acoustic analyses of /jus/ and /juz/ produced at normal and rapid rates revealed that the proportional durations of the semivowel transitions and vowel steady state differed reliable before /s/ and /z/, regardless of total vowel duration. Before /s/ the steady-state duration was approximately 45% that of the entire vowel, as compared with 65% before /z/. These proportional values remained stable across talkers and speaking rates. Postvocalic fricative voicing contrasts apparently may be encoded in the relative timing of articulatory events which produce the entire voiced portion of the syllable. Attunement to both the structure and duration of vowels as sources of perceptual information may then account for the minimal ambiguity of linguistic contrasts assumed to be cued by differences in vowel duration.

Real time digital sound synthesizer

Linearly binary-coded digital control signals representing specific time segments of musical sounds are used in part to control stepped ramp signal generators in a digital synthesizer which runs continuously on a fixed program. The ramp signals are used to control amplitude and frequency parameters of multiple digital oscillators that produce respective constituent tones of the musical sound segments. The synthesizer is operable in response to time-multiplexed digital control signals for multiple musical voices, one voice portion of which is also being computed in real time for multiplexing with previously computed and stored digital control signals for other voices. The indicated computations are effected by known techniques on a commercially available computer to translate performer-actuated transducer output signals, designating notes of a composition, into the aforementioned binary-coded digital control signals.

Acoustic correlates of vocal tension/harshness in the speech of the deaf

Audio recordings were made of twenty young‐adult hearing‐impaired males while reading the first two sentences of the Rainbow Passage. The subjects exhibited a range of severity of vocal tension/harshness as judged from the audio recordings by a panel of three trained listeners. The first ten seconds of the continuous speech samples were digitized and analyzed using an autocorrelation computer program that quantified the periodicity of voiced speech. This measurement compared the power between the periodic portion of the voice (the signal) and the uncorrelated aperiodic component of the wave (the noise) and was expressed as a signal‐to‐noise ratio [J. C. Bancroft, J. Acoust. Soc. Am. Suppl. 1 65, S67(A) (1979)]. The computer program also extracted the average fundamental frequency (F0) for the voiced portion of the signal, as well as the amount of variability of F0. Significant correlations were obtained between the perceptual judgments of severity of vocal tension/harshness and: the signal‐to‐noise ratios; and, the amount of variability in F0. [Work supported by HEW.]