Academia.eduAcademia.edu

Figure 3 – uploaded by Andrey Ronzhin

See full PDFdownloadDownload figure
Figure 3: DET curves for speech/non-speech segment-based classification. Mean and variance of MFCCs are computed over each segment, with (solid line) or without (dashed line) equal-loudness pre-emphasis and cube-root intensity-loudness compression [5]. The minimal values of the corresponding detection cost functions (DCF) are also presented (circles).

Figure 3 DET curves for speech/non-speech segment-based classification. Mean and variance of MFCCs are computed over each segment, with (solid line) or without (dashed line) equal-loudness pre-emphasis and cube-root intensity-loudness compression [5]. The minimal values of the corresponding detection cost functions (DCF) are also presented (circles).

Related Figures (5)

The speech detection method is based on calculation of the information entropy of the signal spectrum as the measure of uncertainty or disorder in a given distribution [6]. The distinction between entropy for speech segments and entropy for background noise is used for speech endpoint detection. Such criterion is less sensitive to the variations of the signal amplitude than the energy- based methods. The method is a modification of the speech detection approach proposed by J.-L. Shen [7] and includes new levels into the analysis of speech signal (Figure 1). The audio signal is divided into short segments with duration 11.6 ms each with overlapping 25%. Short-time signal spectrum is computed using FFT, and normalization of the calculated spectrum over all frequency components is fulfilled giving the probability density function p;. Acceptable values of probability density function are upper and lower bounded. This restriction allows us tc
The speech detection method is based on calculation of the information entropy of the signal spectrum as the measure of uncertainty or disorder in a given distribution [6]. The distinction between entropy for speech segments and entropy for background noise is used for speech endpoint detection. Such criterion is less sensitive to the variations of the signal amplitude than the energy- based methods. The method is a modification of the speech detection approach proposed by J.-L. Shen [7] and includes new levels into the analysis of speech signal (Figure 1). The audio signal is divided into short segments with duration 11.6 ms each with overlapping 25%. Short-time signal spectrum is computed using FFT, and normalization of the calculated spectrum over all frequency components is fulfilled giving the probability density function p;. Acceptable values of probability density function are upper and lower bounded. This restriction allows us tc
where O, and O, are the upper and lower values of probability density, respectively. They hav been experimentally determined to be O, = 0.3 and O, = 0.01. At the next stage the informatio spectral entropy / is estimated, and median smoothing in a window of 5-9 segments is applied Finally, a logical-temporal processing of h (Figure 2) takes into account the possible durations o speech and non-speech fragments.
where O, and O, are the upper and lower values of probability density, respectively. They hav been experimentally determined to be O, = 0.3 and O, = 0.01. At the next stage the informatio spectral entropy / is estimated, and median smoothing in a window of 5-9 segments is applied Finally, a logical-temporal processing of h (Figure 2) takes into account the possible durations o speech and non-speech fragments.
Table 1: Speech/non-speech segment based classification results
Table 1: Speech/non-speech segment based classification results
Table 2: Speech/silence classification results based on spectrum entropy 4. Conclusions In this paper we have applied a two-stage speech detection system. During the first stage, segment- based speech/non-speech classification is performed based on MFCC features and Support Vector Machines within 250 ms accuracy. An improvement is reported if we use loudness equalization and cube-root compression to the power spectrogram after critical-band analysis. Extracted speech segments are further processed through an entropy-based method for speech-endpoint detection within 10 ms accuracy. The proposed system can successfully address the two-fold requirement for robustness and accuracy during the pre-processing stages preceding broadcast speech transcription or speaker diarization.
Table 2: Speech/silence classification results based on spectrum entropy 4. Conclusions In this paper we have applied a two-stage speech detection system. During the first stage, segment- based speech/non-speech classification is performed based on MFCC features and Support Vector Machines within 250 ms accuracy. An improvement is reported if we use loudness equalization and cube-root compression to the power spectrogram after critical-band analysis. Extracted speech segments are further processed through an entropy-based method for speech-endpoint detection within 10 ms accuracy. The proposed system can successfully address the two-fold requirement for robustness and accuracy during the pre-processing stages preceding broadcast speech transcription or speaker diarization.
Related topics:
Speech SegmentationSpeech detectionSpeaker DiarizationBroadcast newsAudio Segmentation
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved