Melodic mode shifting is a construct used occasionally by skilled artists in a raga performance t... more Melodic mode shifting is a construct used occasionally by skilled artists in a raga performance to enhance it by bringing in temporarily shades of a different raga. In this work, we study a specific North Indian Khyal concert structure known as the Jasrangi jugalbandi where a male and female singer co-perform different ragas in an interactive fashion. The mode-shifted ragas with their relatively displaced assumed tonics comprise the identical set of scale intervals and therefore can be easily confused when performed together. With an annotated dataset based on available concerts by well-known artists, we present an analysis of the performance in terms of the raga characteristics as they are manifested through the interactive engagement. We analyse both the aspects of modal music forms, viz. the pitch distribution, representing tonal hierarchy, and the melodic phrases, across the sequence of singing turns by the two artists with reference to representative individual performances of ...
2018 Twenty Fourth National Conference on Communications (NCC), 2018
Onset detection refers to the estimation of the timing of events in a music signal. It is an impo... more Onset detection refers to the estimation of the timing of events in a music signal. It is an important sub-task in music information retrieval and forms the basis of high-level tasks such as beat tracking and tempo estimation. Typically, the onsets of new events in the audio such as melodic notes and percussive strikes are marked by short-time energy rises and changes in spectral distribution. However, each musical instrument is characterized by its own peculiarities and challenges. In this work, we consider the accurate detection of onsets in piano music. An annotated dataset is presented. The operations in a typical onset detection system are considered and modified based on specific observations on the piano music data. In particular, the use of energy-based weighting of multi-band onset detection functions and the use of a new criterion for adapting the final peak-picking threshold are shown to improve the detection of soft onsets in the vicinity of loud notes. We further present a grouping algorithm which reduces spurious onset detections.
Single-channel methods for the separation of the lead vocal from mixed audio have traditionally i... more Single-channel methods for the separation of the lead vocal from mixed audio have traditionally included harmonicsinusoidal modeling and matrix decomposition methods, each with its own strengths and shortcomings. In this work we use a hybrid framework to incorporate prior knowledge about singer and phone identity to achieve the superior separation of the lead vocal from the instrumental background. Singer specific dictionaries learned from available polyphonic recordings provide the soft mask that effectively attenuates the bleeding-through of accompanying melodic instruments typical of purely harmonic-sinusoidal model based separation. The dictionary learning uses NMF optimization across a training set of mixed signal utterances while keeping the vocal signal bases constant across the utterances. A soft mask is determined for each test mixed utterance frame by imposing sparseness constraints in the NMF partial co-factorization. We demonstrate significant improvements in reconstructed signal quality arising from the more accurate estimation of singer-vowel spectral envelope.
A tool for automatic pronunciation evaluation of singing is desirable for those learning a second... more A tool for automatic pronunciation evaluation of singing is desirable for those learning a second language. However, efforts to obtain pronunciation rules for such a tool have been hindered by a lack of data; while many spokenword datasets exist that can be used in developing the tool, there are relatively few sung-lyrics datasets for such a purpose. In this paper, we demonstrate a proof-of-principle for automatic pronunciation evaluation in singing using a knowledge-based approach with limited data in an automatic speech recognition (ASR) framework. To demonstrate our approach, we derive mispronunciation rules specific to SouthEast Asian English accents in singing based on a comparative study of the pronunciation error patterns in singing versus speech. Using training data restricted to American English speech, we evaluate different methods involving the deduced L1-specific (native language) rules for singing. In the absence of L1 phone models, we incorporate the derived pronunciation variations in the ASR framework via a novel approach that combines acoustic models for sub-phonetic segments to represent the missing L1 phones. The word-level assessment achieved by the system on singing and speech is similar, indicating that it is a promising scheme for realizing a full-fledged pronunciation evaluation system for singing in future.
We address the problem of compactly representing the discrete spectral amplitudes of vowel sounds... more We address the problem of compactly representing the discrete spectral amplitudes of vowel sounds produced by a sinusoidal model. A study of frequency warped all pole model representation of spectral amplitudes has been presented. It has been generally accepted that incorporating Bark scale frequency warping in the all-pole modeling improves the perceived accuracy of the modeled sound. However our study suggests that whether such frequency warped all-pole modeling would improve the modeling accuracy depends on the nature of the vowel as well as the voice. We propose an alternative warping function which may be used to improve the modeling accuracy more universally.
I.I.T. Bombay, in collaboration with C.R.L., B.E.L., is working on developing a speech codec to p... more I.I.T. Bombay, in collaboration with C.R.L., B.E.L., is working on developing a speech codec to provide communication quality speech at low bit rates. The codec is required to be robust to acoustic background noise and to channel errors. A sinusoidal coder based on the MBE (Multiband Excitation) model has been adopted as the basic speech representation due to its compact parameter set and the relatively robust, available parameter estimation algorithms. This paper reports a study on the influence of acoustic background noise on the performance of an MBE vocoder with the eventual goal of designing a suitable speech enhancement preprocessor for use in low SNR situations.
Use a parametric representation of audio to train a generative model in the interest of obtaining... more Use a parametric representation of audio to train a generative model in the interest of obtaining more flexible control over the generated sound.
Madam and Prateek. Also I thank Mr. K. S. Nataraj from across the corridor for being a motivation... more Madam and Prateek. Also I thank Mr. K. S. Nataraj from across the corridor for being a motivational speaker. Also I thank my favourite teachers at IITB-Prof. Vikram Gadre, Prof. D. Manjunath, Prof. Pushpak Bhattacharyya and Prof. B. K. Dey-for inspiring and enlightening interactions. Finally, I thank my parents for always supporting me.
The Tabla is one of the principal percussion instruments of the North Indian traditional form of ... more The Tabla is one of the principal percussion instruments of the North Indian traditional form of music. A myriad of sounds can be produced by this instrument, by introducing slight changes in the position and manner of striking. This makes the precise understanding of the idiosyncrasies of every stroke imperative. Previous work has explored the possibility of building a system that can automatically identify, using suitably devised acoustic features, common misarticulations and their nature. In this work, we present a study of the reliability of human perception in deciding tabla stroke quality, in which the extent of agreement among three Tabla experts in ranking strokes of individual tabla bols based on their quality was analyzed. We prepared, for this task, a dataset comprising recordings of individual tabla strokes obtained from Tabla learners of different levels of experience. We also performed a study on the effectiveness of a simple acoustic feature like sustain duration in d...
2017 Twenty-third National Conference on Communications (NCC), 2017
Recordings of read-aloud stories by children in a school setting can be used to provide an assess... more Recordings of read-aloud stories by children in a school setting can be used to provide an assessment of reading skills via automatic speech recognition (ASR). ASR, however, is known to be highly susceptible to background noise. The unusual variety of foreground (breath release, mic pops, etc.) and background (children playing, distinct background talker, wind, etc.) non-speech sounds makes this application particularly challenging. Motivated by the observation on real-world data that close to 50% of the recorded audio comprises purely non-speech activity, we investigate robust approaches to voice activity detection to eliminate non-speech segments to the extent possible prior to ASR. We have exploited energy-based and harmonicity-based features coupled with suitable temporal smoothing constraints in a two-pass noise preprocessing system. A discussion of the voice activity detection performance of the system is presented with reference to the characteristics of the noise types.
The compact representation of harmonic amplitudes in the sinusoidal coding of voiced speech is of... more The compact representation of harmonic amplitudes in the sinusoidal coding of voiced speech is often achieved by the all-pole modeling of a spectral envelope. The perceptual accuracy of the representation may be enhanced by the use of frequency-scale warping prior to LP modeling. The effectiveness of a frequency warping function has been shown in previous work to depend on the underlying signal spectral shape as determined by phoneme quality as well as voice quality. This suggests the possibility of minimizing the perceived degradation due to LP modeling by adapting the warping function to the local signal spectrum. In this paper, an investigation is conducted to determine possible correlations between preferred warping condition and easily computable spectral cues. Experimental results are presented that demonstrate the advantage of adapting the warping function to the signal spectral envelope in the context of a sinusoidal speech coding scheme.
2014 Twentieth National Conference on Communications (NCC), 2014
Voice-based querying for information is a powerful technology that can hugely enhance the scope o... more Voice-based querying for information is a powerful technology that can hugely enhance the scope of information retrieval systems by enabling their remote access via the ubiquitous mobile phone. Information retrieval based on automatic speech recognition however is challenging due to the environment noise and speaker idiosyncrasies typical of real-world scenarios. Wherever possible, strong domain constraints on the language model are used to minimize the impact of signal degradation on recognizer performance. This can lead to grossly mismatched utterance and hypothesis occasionally, a situation that must be detected to protect the call from irrecoverable errors. We consider the revalidation of recognition hypotheses via confidence measures derived from forced alignment using an independent decoder. We show that our FST based aligner can accurately reject incorrect decoder hypotheses while being particularly robust to the phenomenon of incomplete utterances.
2013 International Conference Oriental COCOSDA held jointly with 2013 Conference on Asian Spoken Language Research and Evaluation (O-COCOSDA/CASLRE), 2013
We consider the pronunciation assessment of vowels of Indian English uttered by speakers with Guj... more We consider the pronunciation assessment of vowels of Indian English uttered by speakers with Gujarati L1 using confidence measures obtained by automatic speech recognition. The goodness-of-pronunciation measure as captured by the acoustic likelihood scores can be effective only when the acoustic models used are appropriate for the task i.e. detecting errors in the target language (Indian English) typical of speakers of Gujarati (the source language). Thus the speech data used for acoustic model training is expected to have a prominent influence on system performance. In the absence of labeled speech databases of either the source or target language, we investigate specific combinations of acoustic models trained on available databases of American English and Hindi. It is observed that Indian English speech is better represented by Hindi speech models for vowels common to the two languages rather than by American English models. Further, adaptation with a limited amount of Indian English speech improves the system performance.
The compact representation of harmonic amplitudes in the sinusoidal coding of speech is an import... more The compact representation of harmonic amplitudes in the sinusoidal coding of speech is an important problem in low bit rate speech compression. A widely used method to achieve this is by the all-pole modeling of the spectral envelope. Often a perceptually warped frequency scale is applied in the all-pole modeling to improve perceived accuracy at low model orders. In this work, an attempt is made to obtain a suitable frequency scale warping function by an experimental study on synthetic and natural steady vowels. Subjective listening experiments indicate that the change in perceived quality brought about by frequency warping depends on the underlying signal spectrum or vowel quality. Objective distortion measures are computed to obtain insights into the subjective results. It is observed that an auditory distance measure based on partial loudness shows high correlation with the subjective test scores indicating that frequency masking plays an important role in spectrum envelope modeling.
Traditional short-time spectral attenuation (STSA) speech enhancement algorithms are ineffective ... more Traditional short-time spectral attenuation (STSA) speech enhancement algorithms are ineffective in the presence of highly nonstationary noise due to difficulties in the accurate estimation of the local noise spectrum. With a view to improve the speech quality in the presence of random noise bursts, characteristic of many environmental sounds, a simple postprocessing scheme is proposed that can be applied to the output of an STSA speech enhancement algorithm. The postprocessing algorithm is based on using spectral properties of the noise in order to detect noisy time-frequency regions which are then attenuated using a SNR-based rule. A suitable suppression rule is developed that is applied to the detected noisy regions so as to achieve significant reduction of noise with minimal speech distortion. The post-processing method is evaluated in the context of two well-known STSA speech enhancement algorithms and experimental results demonstrating improved speech quality are presented for a data set of real noise samples.
Unvoiced stops are rapidly varying sounds with acoustic cues to place identity linked to the temp... more Unvoiced stops are rapidly varying sounds with acoustic cues to place identity linked to the temporal dynamics. Neurophysiological studies have indicated the importance of joint spectro-temporal processing in the human perception of stops. In this study, two distinct approaches to modeling the spectro-temporal envelope of unvoiced stop phone segments are investigated with a view to obtaining a low-dimensional feature vector for automatic place classification. Classification accuracies on the TIMIT database and a Marathi words dataset show the overall superiority of classifier combination of polynomial surface coefficients and 2D-DCT. A comparison of performance with published results on the place classification of stops revealed that the proposed spectro-temporal feature systems improve upon the best previous systems' performances. The results indicate that joint spectro-temporal features may be usefully incorporated in hierarchical phone classifiers based on diverse class-specific features.
AbstractThe paper addresses the problem of singing voice detection in Indian Classical Music whe... more AbstractThe paper addresses the problem of singing voice detection in Indian Classical Music where we have the presence of strongly pitched accompaniment. Visual observation of spectra suggests that the temporal fluctuation of the higher harmonics is a strong ...
The Multiband Excitation (MBE) model has been found to be suitable for coding of narrowband speec... more The Multiband Excitation (MBE) model has been found to be suitable for coding of narrowband speech at rates below 4 kbps due to its ability to produce natural sounding speech with a compact set of model parameters and its noise robustness. However the fixed frame analysissynthesis framework of the MBE model is unsuitable for representing dynamic speech sounds such as stops and vocalic transitions. An alternative, the Signal Adaptive MBE model (SA-MBE), is proposed which adapts the time resolution of MBE analysis to the changing characteristics of the speech signal. Based on the transitory nature of speech either 10 or 20 ms analysis resolution is selected. The SA-MBE model is evaluated using subjective and objective methods and found to produce significantly better quality speech with improved intelligibility as compared to the fixed-frame MBE. Preliminary investigations suggest that the SA-MBE model parameters can be quantized to less than 3 kbps.
An investigation of acoustic features relating to vehicular traffic on roadways is reported. Comp... more An investigation of acoustic features relating to vehicular traffic on roadways is reported. Computable features that relate to the type of vehicle and state of motion can be useful in monitoring traffic congestion. In the present work, different vehicles, broadly classified into two, three wheelers and heavy vehicle, are studied for their acoustic signatures. A source filter model of engine sound is used to derive suitable features. The performance of formant based features is compared with that of Mel-frequency cepstral coefficients (MFCC) via a k-NN classifier on a manually labelled database of traffic sounds.
The spectrum interpolation synthesis model has recently been applied in the high quality synthesi... more The spectrum interpolation synthesis model has recently been applied in the high quality synthesis of harmonic musical sounds. In this work we investigate the performance of the model in the compression of music signals. Efficient methods for the automatic analysis, parameter extraction and synthesis of musical signals are presented. The system is tested on several examples of segments from wind and bowed string instruments. It is found that typically a perceived quality matching the original is obtained even when large portions of the waveform are generated by interpolation, implying that a high degree of compression is possible. Further, there is a graceful degradation in quality as the extent of interpolation is increased which makes the model well suited for use in a scalable audio coding framework.
Melodic mode shifting is a construct used occasionally by skilled artists in a raga performance t... more Melodic mode shifting is a construct used occasionally by skilled artists in a raga performance to enhance it by bringing in temporarily shades of a different raga. In this work, we study a specific North Indian Khyal concert structure known as the Jasrangi jugalbandi where a male and female singer co-perform different ragas in an interactive fashion. The mode-shifted ragas with their relatively displaced assumed tonics comprise the identical set of scale intervals and therefore can be easily confused when performed together. With an annotated dataset based on available concerts by well-known artists, we present an analysis of the performance in terms of the raga characteristics as they are manifested through the interactive engagement. We analyse both the aspects of modal music forms, viz. the pitch distribution, representing tonal hierarchy, and the melodic phrases, across the sequence of singing turns by the two artists with reference to representative individual performances of ...
2018 Twenty Fourth National Conference on Communications (NCC), 2018
Onset detection refers to the estimation of the timing of events in a music signal. It is an impo... more Onset detection refers to the estimation of the timing of events in a music signal. It is an important sub-task in music information retrieval and forms the basis of high-level tasks such as beat tracking and tempo estimation. Typically, the onsets of new events in the audio such as melodic notes and percussive strikes are marked by short-time energy rises and changes in spectral distribution. However, each musical instrument is characterized by its own peculiarities and challenges. In this work, we consider the accurate detection of onsets in piano music. An annotated dataset is presented. The operations in a typical onset detection system are considered and modified based on specific observations on the piano music data. In particular, the use of energy-based weighting of multi-band onset detection functions and the use of a new criterion for adapting the final peak-picking threshold are shown to improve the detection of soft onsets in the vicinity of loud notes. We further present a grouping algorithm which reduces spurious onset detections.
Single-channel methods for the separation of the lead vocal from mixed audio have traditionally i... more Single-channel methods for the separation of the lead vocal from mixed audio have traditionally included harmonicsinusoidal modeling and matrix decomposition methods, each with its own strengths and shortcomings. In this work we use a hybrid framework to incorporate prior knowledge about singer and phone identity to achieve the superior separation of the lead vocal from the instrumental background. Singer specific dictionaries learned from available polyphonic recordings provide the soft mask that effectively attenuates the bleeding-through of accompanying melodic instruments typical of purely harmonic-sinusoidal model based separation. The dictionary learning uses NMF optimization across a training set of mixed signal utterances while keeping the vocal signal bases constant across the utterances. A soft mask is determined for each test mixed utterance frame by imposing sparseness constraints in the NMF partial co-factorization. We demonstrate significant improvements in reconstructed signal quality arising from the more accurate estimation of singer-vowel spectral envelope.
A tool for automatic pronunciation evaluation of singing is desirable for those learning a second... more A tool for automatic pronunciation evaluation of singing is desirable for those learning a second language. However, efforts to obtain pronunciation rules for such a tool have been hindered by a lack of data; while many spokenword datasets exist that can be used in developing the tool, there are relatively few sung-lyrics datasets for such a purpose. In this paper, we demonstrate a proof-of-principle for automatic pronunciation evaluation in singing using a knowledge-based approach with limited data in an automatic speech recognition (ASR) framework. To demonstrate our approach, we derive mispronunciation rules specific to SouthEast Asian English accents in singing based on a comparative study of the pronunciation error patterns in singing versus speech. Using training data restricted to American English speech, we evaluate different methods involving the deduced L1-specific (native language) rules for singing. In the absence of L1 phone models, we incorporate the derived pronunciation variations in the ASR framework via a novel approach that combines acoustic models for sub-phonetic segments to represent the missing L1 phones. The word-level assessment achieved by the system on singing and speech is similar, indicating that it is a promising scheme for realizing a full-fledged pronunciation evaluation system for singing in future.
We address the problem of compactly representing the discrete spectral amplitudes of vowel sounds... more We address the problem of compactly representing the discrete spectral amplitudes of vowel sounds produced by a sinusoidal model. A study of frequency warped all pole model representation of spectral amplitudes has been presented. It has been generally accepted that incorporating Bark scale frequency warping in the all-pole modeling improves the perceived accuracy of the modeled sound. However our study suggests that whether such frequency warped all-pole modeling would improve the modeling accuracy depends on the nature of the vowel as well as the voice. We propose an alternative warping function which may be used to improve the modeling accuracy more universally.
I.I.T. Bombay, in collaboration with C.R.L., B.E.L., is working on developing a speech codec to p... more I.I.T. Bombay, in collaboration with C.R.L., B.E.L., is working on developing a speech codec to provide communication quality speech at low bit rates. The codec is required to be robust to acoustic background noise and to channel errors. A sinusoidal coder based on the MBE (Multiband Excitation) model has been adopted as the basic speech representation due to its compact parameter set and the relatively robust, available parameter estimation algorithms. This paper reports a study on the influence of acoustic background noise on the performance of an MBE vocoder with the eventual goal of designing a suitable speech enhancement preprocessor for use in low SNR situations.
Use a parametric representation of audio to train a generative model in the interest of obtaining... more Use a parametric representation of audio to train a generative model in the interest of obtaining more flexible control over the generated sound.
Madam and Prateek. Also I thank Mr. K. S. Nataraj from across the corridor for being a motivation... more Madam and Prateek. Also I thank Mr. K. S. Nataraj from across the corridor for being a motivational speaker. Also I thank my favourite teachers at IITB-Prof. Vikram Gadre, Prof. D. Manjunath, Prof. Pushpak Bhattacharyya and Prof. B. K. Dey-for inspiring and enlightening interactions. Finally, I thank my parents for always supporting me.
The Tabla is one of the principal percussion instruments of the North Indian traditional form of ... more The Tabla is one of the principal percussion instruments of the North Indian traditional form of music. A myriad of sounds can be produced by this instrument, by introducing slight changes in the position and manner of striking. This makes the precise understanding of the idiosyncrasies of every stroke imperative. Previous work has explored the possibility of building a system that can automatically identify, using suitably devised acoustic features, common misarticulations and their nature. In this work, we present a study of the reliability of human perception in deciding tabla stroke quality, in which the extent of agreement among three Tabla experts in ranking strokes of individual tabla bols based on their quality was analyzed. We prepared, for this task, a dataset comprising recordings of individual tabla strokes obtained from Tabla learners of different levels of experience. We also performed a study on the effectiveness of a simple acoustic feature like sustain duration in d...
2017 Twenty-third National Conference on Communications (NCC), 2017
Recordings of read-aloud stories by children in a school setting can be used to provide an assess... more Recordings of read-aloud stories by children in a school setting can be used to provide an assessment of reading skills via automatic speech recognition (ASR). ASR, however, is known to be highly susceptible to background noise. The unusual variety of foreground (breath release, mic pops, etc.) and background (children playing, distinct background talker, wind, etc.) non-speech sounds makes this application particularly challenging. Motivated by the observation on real-world data that close to 50% of the recorded audio comprises purely non-speech activity, we investigate robust approaches to voice activity detection to eliminate non-speech segments to the extent possible prior to ASR. We have exploited energy-based and harmonicity-based features coupled with suitable temporal smoothing constraints in a two-pass noise preprocessing system. A discussion of the voice activity detection performance of the system is presented with reference to the characteristics of the noise types.
The compact representation of harmonic amplitudes in the sinusoidal coding of voiced speech is of... more The compact representation of harmonic amplitudes in the sinusoidal coding of voiced speech is often achieved by the all-pole modeling of a spectral envelope. The perceptual accuracy of the representation may be enhanced by the use of frequency-scale warping prior to LP modeling. The effectiveness of a frequency warping function has been shown in previous work to depend on the underlying signal spectral shape as determined by phoneme quality as well as voice quality. This suggests the possibility of minimizing the perceived degradation due to LP modeling by adapting the warping function to the local signal spectrum. In this paper, an investigation is conducted to determine possible correlations between preferred warping condition and easily computable spectral cues. Experimental results are presented that demonstrate the advantage of adapting the warping function to the signal spectral envelope in the context of a sinusoidal speech coding scheme.
2014 Twentieth National Conference on Communications (NCC), 2014
Voice-based querying for information is a powerful technology that can hugely enhance the scope o... more Voice-based querying for information is a powerful technology that can hugely enhance the scope of information retrieval systems by enabling their remote access via the ubiquitous mobile phone. Information retrieval based on automatic speech recognition however is challenging due to the environment noise and speaker idiosyncrasies typical of real-world scenarios. Wherever possible, strong domain constraints on the language model are used to minimize the impact of signal degradation on recognizer performance. This can lead to grossly mismatched utterance and hypothesis occasionally, a situation that must be detected to protect the call from irrecoverable errors. We consider the revalidation of recognition hypotheses via confidence measures derived from forced alignment using an independent decoder. We show that our FST based aligner can accurately reject incorrect decoder hypotheses while being particularly robust to the phenomenon of incomplete utterances.
2013 International Conference Oriental COCOSDA held jointly with 2013 Conference on Asian Spoken Language Research and Evaluation (O-COCOSDA/CASLRE), 2013
We consider the pronunciation assessment of vowels of Indian English uttered by speakers with Guj... more We consider the pronunciation assessment of vowels of Indian English uttered by speakers with Gujarati L1 using confidence measures obtained by automatic speech recognition. The goodness-of-pronunciation measure as captured by the acoustic likelihood scores can be effective only when the acoustic models used are appropriate for the task i.e. detecting errors in the target language (Indian English) typical of speakers of Gujarati (the source language). Thus the speech data used for acoustic model training is expected to have a prominent influence on system performance. In the absence of labeled speech databases of either the source or target language, we investigate specific combinations of acoustic models trained on available databases of American English and Hindi. It is observed that Indian English speech is better represented by Hindi speech models for vowels common to the two languages rather than by American English models. Further, adaptation with a limited amount of Indian English speech improves the system performance.
The compact representation of harmonic amplitudes in the sinusoidal coding of speech is an import... more The compact representation of harmonic amplitudes in the sinusoidal coding of speech is an important problem in low bit rate speech compression. A widely used method to achieve this is by the all-pole modeling of the spectral envelope. Often a perceptually warped frequency scale is applied in the all-pole modeling to improve perceived accuracy at low model orders. In this work, an attempt is made to obtain a suitable frequency scale warping function by an experimental study on synthetic and natural steady vowels. Subjective listening experiments indicate that the change in perceived quality brought about by frequency warping depends on the underlying signal spectrum or vowel quality. Objective distortion measures are computed to obtain insights into the subjective results. It is observed that an auditory distance measure based on partial loudness shows high correlation with the subjective test scores indicating that frequency masking plays an important role in spectrum envelope modeling.
Traditional short-time spectral attenuation (STSA) speech enhancement algorithms are ineffective ... more Traditional short-time spectral attenuation (STSA) speech enhancement algorithms are ineffective in the presence of highly nonstationary noise due to difficulties in the accurate estimation of the local noise spectrum. With a view to improve the speech quality in the presence of random noise bursts, characteristic of many environmental sounds, a simple postprocessing scheme is proposed that can be applied to the output of an STSA speech enhancement algorithm. The postprocessing algorithm is based on using spectral properties of the noise in order to detect noisy time-frequency regions which are then attenuated using a SNR-based rule. A suitable suppression rule is developed that is applied to the detected noisy regions so as to achieve significant reduction of noise with minimal speech distortion. The post-processing method is evaluated in the context of two well-known STSA speech enhancement algorithms and experimental results demonstrating improved speech quality are presented for a data set of real noise samples.
Unvoiced stops are rapidly varying sounds with acoustic cues to place identity linked to the temp... more Unvoiced stops are rapidly varying sounds with acoustic cues to place identity linked to the temporal dynamics. Neurophysiological studies have indicated the importance of joint spectro-temporal processing in the human perception of stops. In this study, two distinct approaches to modeling the spectro-temporal envelope of unvoiced stop phone segments are investigated with a view to obtaining a low-dimensional feature vector for automatic place classification. Classification accuracies on the TIMIT database and a Marathi words dataset show the overall superiority of classifier combination of polynomial surface coefficients and 2D-DCT. A comparison of performance with published results on the place classification of stops revealed that the proposed spectro-temporal feature systems improve upon the best previous systems' performances. The results indicate that joint spectro-temporal features may be usefully incorporated in hierarchical phone classifiers based on diverse class-specific features.
AbstractThe paper addresses the problem of singing voice detection in Indian Classical Music whe... more AbstractThe paper addresses the problem of singing voice detection in Indian Classical Music where we have the presence of strongly pitched accompaniment. Visual observation of spectra suggests that the temporal fluctuation of the higher harmonics is a strong ...
The Multiband Excitation (MBE) model has been found to be suitable for coding of narrowband speec... more The Multiband Excitation (MBE) model has been found to be suitable for coding of narrowband speech at rates below 4 kbps due to its ability to produce natural sounding speech with a compact set of model parameters and its noise robustness. However the fixed frame analysissynthesis framework of the MBE model is unsuitable for representing dynamic speech sounds such as stops and vocalic transitions. An alternative, the Signal Adaptive MBE model (SA-MBE), is proposed which adapts the time resolution of MBE analysis to the changing characteristics of the speech signal. Based on the transitory nature of speech either 10 or 20 ms analysis resolution is selected. The SA-MBE model is evaluated using subjective and objective methods and found to produce significantly better quality speech with improved intelligibility as compared to the fixed-frame MBE. Preliminary investigations suggest that the SA-MBE model parameters can be quantized to less than 3 kbps.
An investigation of acoustic features relating to vehicular traffic on roadways is reported. Comp... more An investigation of acoustic features relating to vehicular traffic on roadways is reported. Computable features that relate to the type of vehicle and state of motion can be useful in monitoring traffic congestion. In the present work, different vehicles, broadly classified into two, three wheelers and heavy vehicle, are studied for their acoustic signatures. A source filter model of engine sound is used to derive suitable features. The performance of formant based features is compared with that of Mel-frequency cepstral coefficients (MFCC) via a k-NN classifier on a manually labelled database of traffic sounds.
The spectrum interpolation synthesis model has recently been applied in the high quality synthesi... more The spectrum interpolation synthesis model has recently been applied in the high quality synthesis of harmonic musical sounds. In this work we investigate the performance of the model in the compression of music signals. Efficient methods for the automatic analysis, parameter extraction and synthesis of musical signals are presented. The system is tested on several examples of segments from wind and bowed string instruments. It is found that typically a perceived quality matching the original is obtained even when large portions of the waveform are generated by interpolation, implying that a high degree of compression is possible. Further, there is a graceful degradation in quality as the extent of interpolation is increased which makes the model well suited for use in a scalable audio coding framework.
Uploads
Papers by Preeti Rao