Dnsmos: A Non-Intrusive Perceptual Objective Speech Quality Metric to Evaluate Noise Suppressors

Reddy, Chandan K.; Gopal, Vishak; Cutler, Ross

doi:10.1109/icassp39728.2021.9414878

Cited by 117 publications

(87 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The speech recognition system was trained to handle audio with a wide range of energy levels so we do not expect any degradation of WAcc due to varying energy levels in clips from the blind test set. Table 1 shows the Pearson correlation coefficient (PCC) and Spearman's rank correlation coefficient (SRCC) between per-model subjective scores and corresponding DNSMOS P.835 scores [9]. The high correlation between subjective scores and DNSMOS P.835 in both tracks shows the efficacy of DNSMOS P.835 in ranking the DNS models.…”

Section: Resultsmentioning

confidence: 99%

“…We provided the participants with an Azure API for estimating WAcc on the development set. We computed DNSMOS P.835 [9] for each audio clip in the training set and provided this to participants. DNSMOS P.835 scores can be used to segment the training dataset for conducting the data ablation studies.…”

Section: Challenge Tracksmentioning

confidence: 99%

“…We provide fullband datasets for training personalized and non-personalized DNS models. We collected real-world test sets, developed a framework for P.835 subjective evaluation, created APIs for DNSMOS P.835 [9] and WAcc. We provided the development test set, DNSMOS P.835 API, and WAcc API at the beginning of the challenge, which helps participants to optimize their models.…”

Section: Introductionmentioning

confidence: 99%

“…We evaluated the submitted models based on ITU-T P.835 subjective evaluation scores, namely speech quality (SIG), background noise quality (BAK), and overall audio quality (OVRL), as well as WAcc from a state-of-the-art speech recognition system. In addition, we make DNSMOS P.835 [9] freely available to researchers. DNS-MOS P.835 is a deep learning model that predicts SIG, BAK, OVRL scores for a noisy test clip.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

ICASSP 2022 Deep Noise Suppression Challenge

Dubey¹,

Gopal²,

Cutler³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

The Deep Noise Suppression (DNS) challenge is designed to foster innovation in the area of noise suppression to achieve superior perceptual speech quality. This is the 4th DNS challenge, with the previous editions held at INTERSPEECH 2020 [1], ICASSP 2021 [2], and INTERSPEECH 2021 [3]. We open-source datasets and test sets for researchers to train their deep noise suppression models, as well as a subjective evaluation framework based on ITU-T P.835 to rate and rank-order the challenge entries. We provide access to DNS-MOS P.835 and word accuracy (WAcc) APIs to challenge participants to help with iterative model improvements. In this challenge, we introduced the following changes: (i) Included mobile device scenarios in the blind test set; (ii) Included a personalized noise suppression track with baseline; (iii) Added WAcc as an objective metric; (iv) Included DNSMOS P.835; (v) Made the training datasets and test sets fullband (48 kHz). We use an average of WAcc and subjective scores P.835 SIG, BAK, and OVRL to get the final score for ranking the DNS models. We believe that as a research community, we still have a long way to go in achieving excellent speech quality in challenging noisy real-world scenarios.

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Challenge Tracksmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

ICASSP 2022 Deep Noise Suppression Challenge

Dubey¹,

Gopal²,

Cutler³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…We use DNSMOS [19] which is a reliable non-intrusive objective speech quality metric as our evaluation metrics at training stag and take Mean Opinion Score (MOS) of the ITU-T P.835 framework as result. The results of the evaluation using the ITU-T P.835 criterion [20] which is provided by the organizer are shown in Table 2.…”

Section: Deep Noise Suppression Challengementioning

confidence: 99%

DBNet: A Dual-branch Network Architecture Processing on Spectrum and Waveform for Single-channel Speech Enhancement

Zhang¹,

et al. 2021

Preprint

View full text Add to dashboard Cite

In real acoustic environment, speech enhancement is an arduous task to improve the quality and intelligibility of speech interfered by background noise and reverberation. Over the past years, deep learning has shown great potential on speech enhancement. In this paper, we propose a novel real-time framework called DBNet which is a dual-branch structure with alternate interconnection. Each branch incorporates an encoderdecoder architecture with skip connections. The two branches are responsible for spectrum and waveform modeling, respectively. A bridge layer is adopted to exchange information between the two branches. Systematic evaluation and comparison show that the proposed system substantially outperforms related algorithms under very challenging environments. And in INTERSPEECH 2021 Deep Noise Suppression (DNS) challenge, the proposed system ranks the top 8 in real-time track 1 in terms of the Mean Opinion Score (MOS) of the ITU-T P.835 framework.

show abstract

Dual-Channel Speech Enhancement Using Neural Network Adaptive Beamforming

Jiang

Liu

Shuai

et al. 2022

Communications and Networking

View full text Add to dashboard Cite

Dual-channel speech enhancement based on traditional beamforming is difficult to effectively suppress noise. In recent years, it is promising to replace beamforming with a neural network that learns spectral characteristic. This paper proposes a neural network adaptive beamforming end-to-end dual-channel model for speech enhancement task. First, the LSTM layer is used to directly process the original speech waveform to estimate the time-domain beamforming filter coefficients of each channel and convolve and sum it with the input speech. Second, we modified a fully-convolutional time-domain audio separation network (Conv-TasNet) into a network suitable for speech enhancement which is called Denoising-TasNet to further enhance the output of the beamforming. The experimental results show that the proposed method is better than convolutional recurrent network (CRN) model and several popular noise reduction methods.

show abstract

Dnsmos: A Non-Intrusive Perceptual Objective Speech Quality Metric to Evaluate Noise Suppressors

Cited by 117 publications

References 12 publications

ICASSP 2022 Deep Noise Suppression Challenge

ICASSP 2022 Deep Noise Suppression Challenge

DBNet: A Dual-branch Network Architecture Processing on Spectrum and Waveform for Single-channel Speech Enhancement

Dual-Channel Speech Enhancement Using Neural Network Adaptive Beamforming

Contact Info

Product

Resources

About