Robust excitation-based features for Automatic Speech Recognition

Drugman, Thomas; Stylianou, Yannis; Chen, Langzhou; Xie, Chen; Gales, Mark J. F.

doi:10.1109/icassp.2015.7178855

Cited by 6 publications

(4 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Furthermore, the improvements in ASR performance are consistently seen across all the noisy test conditions and with a sophisticated RNN-LM. In addition, the performance achieved is also considerably better than the results such as excitation based features (EB) reported by [33].…”

Section: Experiments and Resultsmentioning

confidence: 62%

“…For each dataset, we compare the ASR performance of the proposed approach of learning acoustic representation from raw waveform with acoustic FB (A) with relevance weighting (A-R) and modulation FB (M) with relevance weighting (M-R) denoted as (A-R,M-R), traditional log mel filterbank energy (MFB) features (80 dimension), power normalized filterbank energy (PFB) features [31], mean Hilbert envelope (MHE) features [32], and excitation based (EB) features [33]. We also compare performance with the SincNet method proposed in [11].…”

Section: Experiments and Resultsmentioning

confidence: 99%

See 1 more Smart Citation

Interpretable Representation Learning for Speech and Audio Signals Based on Relevance Weighting

Agrawal

Ganapathy

2020

IEEE/ACM Trans. Audio Speech Lang. Process.

View full text Add to dashboard Cite

The learning of interpretable representations from raw data presents significant challenges for time series data like speech. In this work, we propose a relevance weighting scheme that allows the interpretation of the speech representations during the forward propagation of the model itself. The relevance weighting is achieved using a sub-network approach that performs the task of feature selection. A relevance subnetwork, applied on the output of first layer of a convolutional neural network model operating on raw speech signals, acts as an acoustic filterbank (FB) layer with relevance weighting. A similar relevance sub-network applied on the second convolutional layer performs modulation filterbank learning with relevance weighting. The full acoustic model consisting of relevance subnetworks, convolutional layers and feed-forward layers is trained for a speech recognition task on noisy and reverberant speech in the Aurora-4, CHiME-3 and VOiCES datasets. The proposed representation learning framework is also applied for the task of sound classification in the UrbanSound8K dataset. A detailed analysis of the relevance weights learned by the model reveals that the relevance weights capture information regarding the underlying speech/audio content. In addition, speech recognition and sound classification experiments reveal that the incorporation of relevance weighting in the neural network architecture improves the performance significantly.

show abstract

Section: Experiments and Resultsmentioning

confidence: 62%

Section: Experiments and Resultsmentioning

confidence: 99%

Interpretable Representation Learning for Speech and Audio Signals Based on Relevance Weighting

Agrawal

Ganapathy

2020

IEEE/ACM Trans. Audio Speech Lang. Process.

View full text Add to dashboard Cite

show abstract

“…At the same time it is also noticed that under certain noise conditions, while vocal tract information is lost, fundamental frequency information is still retained. Pitch or fundamental frequency is an excitation source information and excitation source features are robust to noisy conditions [4]. In a study conducted on Japanese Newspaper Article Sentences database, it was noticed that inclusion of F0 information resulted in an absolute improvement of about 9% in noisy speech recognition [5].…”

Section: Introductionmentioning

confidence: 99%

Robust Mizo digit recognition using data augmentation and tonal information

Sarma¹,

Dey²,

Lalhminghlui³

et al. 2018

Speech Prosody 2018

View full text Add to dashboard Cite

Performance of speech recognition system severely degrades in noisy environment. Considering this, in this work, we present a method to improve performance of a Mizo digit recognition system in different noisy conditions using data augmentation and tonal information. Mizo is a tonal language and each digit in Mizo is spoken with one of the four tones present in the language. Therefore, the tone contains information about the spoken digit. Tone is related to the excitation source and excitation source information is robust to noisy conditions when compared with the vocal tract information. Normalized cross correlation function, pitch and pitch dynamics are used as additional features to represent the tonal information and improvement is achieved in Mel frequency cepstral coefficient (MFCC) based baseline systems in noisy conditions. Data augmentation is another technique used in the literature for robust speech recognition. Use of data augmentation further improves the performance of the Mizo digit recognition.

show abstract

“…In case of noisy speech conditions, the vocal tract information gets severely affected, yet the fundamental frequency (F0) information is largely retained. Drugman et al [8] reported that The data involved in this tonal analysis is described in [10,11].…”

Section: Introductionmentioning

confidence: 99%

Robust Mizo Continuous Speech Recognition

et al. 2018

View full text Add to dashboard Cite

Mizo is an under-resourced tonal language that is mainly spoken in NorthEast India. It has 4 canonical tones along with a tone-sandhi. In Mizo language, a majority of the words contain tone information. As a result of that, it exhibits higher acoustic variability like other tonal languages in the world. In this work, we investigate the impact of tonal information on robust Mizo continuous speech recognition (CSR). First, separate baseline CSR systems are developed employing the Mel-frequency cepstral coefficient (MFCC) based acoustic features and salient acoustic modeling paradigms. For further improvement, the tonal information has been incorporated in each of the CSR systems. For this purpose, 3-dimensional tonal features are derived which include pitch, pitch-difference, and probability of voicing values. Our experimental study reveals that with the inclusion of tonal information, the robustness of Mizo CSR system gets enhanced across all acoustic modeling paradigms. This trend is attributed to lesser degradation in the fundamental frequency information than the vocal tract information under noisy conditions.

show abstract

Robust excitation-based features for Automatic Speech Recognition

Cited by 6 publications

References 32 publications

Interpretable Representation Learning for Speech and Audio Signals Based on Relevance Weighting

Interpretable Representation Learning for Speech and Audio Signals Based on Relevance Weighting

Robust Mizo digit recognition using data augmentation and tonal information

Robust Mizo Continuous Speech Recognition

Contact Info

Product

Resources

About