Clipped Speech Research Articles

Monitoring Voip Speech Quality for Chopped and Clipped Speech

Real-time monitoring of speech quality for VoIP calls is a significant challenge. This paper presents early work on a no-reference objective model for quantifying perceived speech quality in VoIP. The overall approach uses a modular design that will be able to help pinpoint the reason for degradations as well as quantifying their impact on speech quality. The model is being designed to work with narrowband and wideband signals. This initial work is focused on rating amplitude clipped or chopped speech, which are common problems in VoIP. A model sensitive to each of these degradations is presented and then tested with both synthetic and real examples of chopped and clipped speech. The results were compared with predicted MOS outputs from four objective speech quality models: ViSQOL, PESQ, POLQA and P.563. The model output showed consistent relationships between this model's clip and chop detection modules and the quality predictions from the other objective speech quality models. Further work is planned to widen the range of degradation types captured by the model, such as non-stationary background noise and speaker echo. While other components (e.g. a voice activity detector) would be necessary to deploy the model for stand-alone VoIP monitoring, the results show good potential for using the model in a realtime monitoring tool.

Detection and reconstruction of clipped speech for speaker recognition

Clipping is often observed in speech acquisition, due to the limited numerical range or the non-linear compensation of recording devices. The clipping inevitably changes the spectrum of speech signals, and thus partially distorts the speaker information contained in the signal. This paper investigates the impact of signal clipping on speaker recognition, and proposes a simple yet effective clipping detection approach as well as a signal reconstruction approach based on deep neural networks (DNNs). The experiments are conducted on the core test of the NIST SRE2008 task by simulating clipped speech at various clipping rates. The results show that clipping does impact the performance of speaker recognition, but the impact is rather marginal unless the clipping rate is larger than 80%. We also find that the simple distribution-based detection method is capable of detecting clipped speech with a higher accuracy, and the DNN-based reconstruction can achieve promising performance gains for speaker recognition on clipped speech.

Restoring clipped speech signal based on spectral transformation of each frequency band

In recent years, high-quality speech recording is requested for comfortably using a communication system. However, a clipping distortion caused by input exceeding the maximum range of amplifier is one of the problems with a sound quality degradation. Although linear prediction method has been conventionally proposed for restoring a clipped speech signal, it has a problem that the frequently clipped speech signal degrades the restoration performance by increasing the prediction errors. In this paper, we propose a method for restoring a clipped speech signal based on a spectral transformation of each frequency band. In this method, the spectral envelope of target speech signal in each frame is approximated to the spectral envelope of original speech signal to remove the influence of a clipping distortion. In particular, the spectral envelope in higher frequency domain including a static characteristic of the speaker is replaced with the spectral envelope of the unclipped speech signal prepared in advance. Then, the spectral envelope in lower frequency domain including a characteristic of phoneme is approximated with Gaussian Mixture Models. We carried out an evaluation experiment for sound quality of speech signal processed by the proposed method. As a result, we confirmed the effectiveness of the proposed method.

An Asymmetric Speech Coding Algorithm Using Vector Quantization and Silence Deletion

AbstractIn store‐and‐forward and multi‐media speech applications, e.g. in voice‐mail or when adding verbal comments to a document, the speech encoder can be allowed more time and higher complexity than in real‐time speech coding. Decoding however must be simple when a low‐cost terminal (e.g., a PC) is used. This kind of coding algorithm is called an asymmetric algorithm. Asymmetric algorithms exploit encoding delay to code the signal efficiently. An existing asymmetric coding algorithm is Vector Quantization (VQ). A new asymmetric algorithm is proposed, based on the coding of the time intervals between zero‐crossings (frames) using VQ and silence deletion. The algorithm takes advantage of the empirical fact that infinitely clipped speech retains its intelligibility. The algorithm has been implemented using software and its performance was evaluated.

Clipped speech‐linear predictive coding speech processor

The present invention relates to a speech recognition system and the method therefor, which analyzes a sampled clipped speech signal for identifying a spoken utterance. An input signal representative of the spoken utterance is passed through a clipper to generate a clipped input signal. A sampler generates a plurality of discrete binary values, each discrete binary value corresponding to a sample value of the clipped input signal. A processor then analyzes the plurality of sample values thereby identifying the spoken utterance. Analysis includes determining linear prediction coefficients of the autocorrelation function of speech utterences.

Cues in the perception of infinitely clipped speech

“Infinitely clipped” (“IC”) speech is surprisingly intelligible [Licklider and Pollack, J. Acoust. Soc. Am. 20, 42–51 (1948)]. Because infinite clipping so severely distorts the speech spectrum, it is sometimes speculated that a time‐domain similarity between natural and IC speech—precise identity of zero‐crossing locations—might account for the latter's intelligibility. This hypothesis has practical implications (should speech recognition algorithms look beyond power spectra?) and theoretical ones [cf. Scott, J. Acoust. Soc. Am. 60, 1354–1365 (1976)]. However, the hypothesis would be refuted if it turned out that the small amount of residual similarity between the power spectra of the original and IC signals could alone account for the observed intelligibility. Three types of stimuli—natural vowels, /p,t,k/ in natural‐sentence context, and synthetic vowels—were infinitely clipped (“IC stimuli”), then subjected to transformations which preserved the power spectrum but grossly distorted the phase spectrum, thereby destroying the original zero‐crossing information. These transformed IC stimuli were no less intelligible than the unaltered IC stimuli, strongly suggesting that power spectrum rather than zero‐crossing cues are responsible for IC‐speech intelligibility. Further, constant‐parameter, synthetic‐vowel inteiligibility decreased dramatically upon clipping, natural‐vowel intelligibility relatively little, demonstrating the usefulness of spectral change and/or prosodic cues in the absence of accurate formant information.

Multichannel infinite clipping as a form of sampling of speech signals

A remarkable improvement of both intelligibility and naturalness of infinitely clipped speech can be achieved by means of a multichannel system in which the speech signal is splitted into several band-pass channels before the clipping and successively reconstructed by summing up the clipped outputs of each channel.

Perception of (Hindi) consonants in differentiated clipped speech.

Twenty nine initial consonants and 31 final consonants of CVC syllables were studied for perceptual confusions under the condition of differentialtion and infinite peak clipping. A quantitative procedure was used to calculate the percentage of information of each of the nine linguistic features (affriction, friction, nasality, aspiration, voicing, liquids, flapped liquids, continuants, and place), and the data so obtained were studied in relation to the data of the two previous studies [J. Acoust. Soc. Am. 45, 758–763 (1969) and J. Acoust. Soc. Am. 45, 770–773 (1969)]. It was found that liquids, voicing and continuants were more affected and that friction, affrication and place of articulation were less affected by the distortion of differentiation and clipping. The relative importance of features in differentiated clipped speech was not found to be the same as that of the normal speech. The study also indicated the perceptual difference between the two distortion conditions with and without pre differentiation. Subject Classification: 70.30.

Consonantal Context and Vowel Perception in Normal and Clipped Speech

An experiment was done to study the effect of consonantal context on vowel perception in normal and infinitely clipped speech. In this experiment, 10 Hindi vowels | i, e, ∈, a, Λ, ɔ, o, U, u | were placed in | hΛj—J | and |hΛw—w| contexts, and recorded by two male speakers on a tape. The recordings were presented to six well-trained listeners with or without infinite peak clipping. Listeners' responses were summarized in a confusion matrix for each context and for each condition of testing. Thus, four confusion matrices were formed. Study of the confusion matrices reveals that the consonantal context does affect the perception of vowels in some of its aspects. Vowel intelligibility was found to be very different between the two contexts, while the relative intelligibility of vowels and specific confusions among vowels were not found to be significantly different.

Intelligibility Cues in Normal and Distorted Speech Forms

Normal and distorted but high-intelligibility speech forms are compared using a millisecond-resolution analog ear display. Pattern details which convey intelligibility are identified separately from those which relate to matters of quality. Distorted speech forms include high-pass filtered normal speech, whispered speech, whistled speech, helium speech, and infinitely clipped speech. The concept of a single principal formant is partly justified. In addition, there exist brief periods when activity penetrates into the ear further than the position of the formant, these events being periodic or random for voiced or whispered sounds, respectively.

Interpretation of infinitely clipped speech properties

The paper is based upon the statistical theory of signals. It is shown that the variation of the clipped Gaussian noise spectrum is 5 percent, while the normalized cross-correlation function is 0.80; the results are extended to more general cases. A property of the conditional mean value of speech signals is derived. A typical value of the normalized intelligibility of clipped speech in the presence of noise is 0.70 for S/N = 3 dB; assuming the ear is a clipper, theoretical results are well in agreement. The existence of an intelligibility threshold is shown. The need for information on the envelope (or dynamics) of the speech wave to meet the requirement of naturalness is underlined.

Theoretical Investigations to Reduce Harmonic Distortions in a Clipping Process

To investigate harmonic distortions of clipped speech, the speech waves are replaced by a sum of two sine waves. Infinite clipping of these two frequencies results in a harmonic distortion of 75%. If the audioband is modulated to a higher frequency range, one of the sidebands clipped, and all distortions beyond the single band width suppressed, the distortion drops to 50%. Distortion is further reduced if the original representative sum of the sine waves is differentiated and transformed to a pulse wave with pulse width modulation, the width corresponding to amplitude maximums and minimums. With pulse width modulation of unit amplitude, the advantages of normal clipped speech transmission are retained. On the receiver side the pulse width modulation is transformed to a pulse amplitude modulation with relatively short pulses. The pulse wave, still highly distorted is applied to a storage device (e.g., a magnetic tape or a ladder network of condensers). The picture of the pulses is smoothed and approximates the picture of the original sum of the two sine waves. Essentially, the approximation is the replacement of the original wave by a triangular one. This process, mathematically seen as a sine modulation of an unit triangular curve, reduced harmonic distortions to 12% instead of 75% as in normal infinite clipping.