Motivated by the problem of improving the performance of speech enhancement algorithms in non-stationary acoustic environments with low SNR, a framework is proposed for identifying signal frames of noisy speech that are unlikely to contain voice activity. Such voice-inactive frames can then be incorporated into an adaptation strategy to improve the performance of existing speech enhancement algorithms. This adaptive approach is applicable to single-channel as well as multi-channel algorithms for noisy speech. In both cases, the adaptive versions of the enhancement algorithms are observed to improve SNR levels by 20dB, as indicated by PESQ and WER criteria.
In advanced speech enhancement algorithms, it is often of interest to identify some regions of the signal that have a high likelihood of being noise only i.e. no speech present. This is in contrast to advanced speech recognition, speaker recognition, and pitch tracking algorithms in which we are interested in identifying all regions that have a high likelihood of containing speech, as well as regions that have a high likelihood of not containing speech. In other terms, this would mean minimizing the false positive and false negative rates, respectively. In the context of speech enhancement, the identification of some speech-absent regions prompts the minimization of false positives while setting an acceptable tolerance on false negatives, as determined by the performance of the enhancement algorithm.
Typically, Voice Activity Detectors (VADs) are used for identifying speech absent regions for the application of speech enhancement. In recent years a myriad of Deep Neural Network (DNN) based approaches have been proposed to improve the performance of VADs at low SNR levels by training on combinations of speech and noise. Training on such an exhaustive dataset is combinatorically explosive. For this dissertation, we propose a voice inactivity ranking framework, where the identification of voice-inactive frames is performed using a machine learning (ML) approach that only uses clean speech utterances for training and is robust to high levels of noise. In the proposed framework, input frames of noisy speech are ranked by ‘voice inactivity score’ to acquire definitely speech inactive (DSI) frame-sequences. These DSI regions serve as a noise estimate and are adaptively used by the underlying speech enhancement algorithm to enhance speech from a speech mixture.
The proposed voice-inactivity ranking framework was used to perform speech enhancement in single-channel and multi-channel systems. In the context of microphone arrays, the proposed framework was used to determine parameters for spatial filtering using adaptive beamformers. We achieved an average Word Error Rate (WER) improvement of 50% at SNR levels below 0dB compared to the noisy signal, which is 7±2.5% more than the framework where state-of-the-art VAD decision was used for spatial filtering. For monaural signals, we propose a multi-frame multiband spectral-subtraction (MF-MBSS) speech enhancement system utilizing the voice inactivity framework to compute and update the noise statistics on overlapping frequency bands. The proposed MF-MBSS not only achieved an average PESQ improvement of 16% with a maximum improvement of 56% when compared to the state-of-the-art Spectral Subtraction but also a 5 ± 1.5% improvement in the Word Error Rate (WER) of the spatially filtered output signal, in non-stationary acoustic environments.
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/43714 |
| Date | 26 January 2022 |
| Creators | Sharma, Siddhant |
| Contributors | Nawab, S. Hamid |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
Page generated in 0.0025 seconds