The quality of experience (QoE) delivered by video conferencing systems is significantly influenced by accurately estimating the time-varying available bandwidth between the sender and receiver. Bandwidth estimation for real-time communications remains an open challenge due to rapidly evolving network architectures, increasingly complex protocol stacks, and the difficulty of defining QoE metrics that reliably improve user experience. In this work, we propose a deployed, human-in-the-loop, data-driven framework for bandwidth estimation to address these challenges. Our approach begins with training objective QoE reward models derived from subjective user evaluations to measure audio and video quality in real-time video conferencing systems. Subsequently, we collect roughly 1M network traces with objective QoE rewards from real-world Microsoft Teams calls to curate a bandwidth estimation training dataset. We then introduce a novel distributional offline reinforcement learning (RL) algorithm to train a neural-network-based bandwidth estimator aimed at improving QoE for users. Our real-world A/B test demonstrates that the proposed approach reduces the subjective poor call ratio by 11.41% compared to the baseline bandwidth estimator. Furthermore, the proposed offline RL algorithm is benchmarked on D4RL tasks to demonstrate its generalization beyond bandwidth estimation.

Real-time video applications require accurate bandwidth estimation (BWE) to maintain user experience across varying network conditions. However, increasing network heterogeneity challenges general-purpose BWE algorithms, necessitating solutions that adapt to end-user environments. While widely adopted, heuristic-based methods are difficult to individualize without extensive domain expertise. Conversely, online reinforcement learning (RL) offers ease of customization but neglects prior domain expertise and suffers from sample inefficiency. Thus, we present Merlin, an imitation learning-based solution that replaces the manual parameter tuning of heuristic-based methods with data-driven updates to streamline end-user personalization. Our key insight is that transforming heuristic-based BWE algorithms into neural networks facilitates data-driven personalization. Merlin utilizes Behavioral Cloning to efficiently learn from offline telemetry logs, capturing heuristic policies without live network interactions. The cloned policy can then be seamlessly tailored to end user network conditions through online finetuning. In real intercontinental videoconferencing calls, Merlin matches our heuristic's policy with no statistically significant differences in user quality of experience (QoE). Finetuning Merlin's control policy to end-user environments enables QoE improvements of up to 7.8 % compared to the heuristic policy. Lastly, our IL-based design performs competitively with current state-of-the-art online RL techniques but converges with 80 % fewer videoconferencing samples, facilitating practical end-user personalization.

Photorealistic avatars are human avatars that look, move, and talk like real people. The performance of photorealistic avatars has significantly improved recently based on objective metrics such as PSNR, SSIM, LPIPS, FID, and FVD. However, recent photorealistic avatar publications do not provide subjective tests of the avatars to measure human usability factors. We provide an open source test framework to subjectively measure photorealistic avatar performance in ten dimensions: realism, trust, comfortableness using, comfortableness interacting with, appropriateness for work, creepiness, formality, affinity, resemblance to the person, and emotion accuracy. Using telecommunication scenarios, we show that the correlation of nine of these subjective metrics with PSNR, SSIM, LPIPS, FID, and FVD is weak, and moderate for emotion accuracy. The crowdsourced subjective test framework is highly reproducible and accurate when compared to a panel of experts. We analyze a wide range of avatars from photorealistic to cartoon-like and show that some photorealistic avatars are approaching real video performance based on these dimensions. We also find that for avatars above a certain level of realism, eight of these measured dimensions are strongly correlated. This means that avatars that are not as realistic as real video will have lower trust, comfortableness using, comfortableness interacting with, appropriateness for work, formality, and affinity, and higher creepiness compared to real video. In addition, because there is a strong linear relationship between avatar affinity and realism, there is no uncanny valley effect for photorealistic avatars in the telecommunication scenario. We suggest several extensions of this test framework for future work and discuss design implications for telecommunication systems. The test framework is available at https://github.com/microsoft/P.910.

We introduce a distinctive real-time, causal, neural network-based active speaker detection system optimized for low-power edge computing. This system drives a virtual cinematography module and is deployed on a commercial device. The system uses data originating from a microphone array and a 360-degree camera. Our network requires only 127 MFLOPs per participant, for a meeting with 14 participants. Unlike previous work, we examine the error rate of our network when the computational budget is exhausted, and find that it exhibits graceful degradation, allowing the system to operate reasonably well even in this case. Departing from conventional DOA estimation approaches, our network learns to query the available acoustic data, considering the detected head locations. We train and evaluate our algorithm on a realistic meetings dataset featuring up to 14 participants in the same meeting, overlapped speech, and other challenging scenarios.

The quality of experience (QoE) delivered by video conferencing systems to end users depends in part on correctly estimating the capacity of the bottleneck link between the sender and the receiver over time. Bandwidth estimation for real-time communications (RTC) remains a significant challenge, primarily due to the continuously evolving heterogeneous network architectures and technologies. From the first bandwidth estimation challenge which was hosted at ACM MMSys 2021, we learned that bandwidth estimation models trained with reinforcement learning (RL) in simulations to maximize network-based reward functions may not be optimal in reality due to the sim-to-real gap and the difficulty of aligning network-based rewards with user-perceived QoE. This grand challenge aims to advance bandwidth estimation model design by aligning reward maximization with user-perceived QoE optimization using offline RL and a real-world dataset with objective rewards which have high correlations with subjective audio/video quality in Microsoft Teams. All models submitted to the grand challenge underwent initial evaluation on our emulation platform. For a comprehensive evaluation under diverse network conditions with temporal fluctuations, top models were further evaluated on our geographically distributed testbed by using each model to conduct 600 calls within a 12-day period. The winning model is shown to deliver comparable performance to the top behavior policy in the released dataset. By leveraging real-world data and integrating objective audio/video quality scores as rewards, offline RL can therefore facilitate the development of competitive bandwidth estimators for RTC.

Current state-of-the-art video quality models, such as VMAF, give excellent prediction results by comparing the degraded video with its reference video. However, they do not consider temporal distortions (e.g., frame freezes or skips) that occur during videoconferencing calls. In this paper, we present a data-driven approach for modeling such distortions automatically by training an LSTM with subjective quality ratings labeled via crowdsourcing. The videos were collected from live videoconferencing calls in 83 different network conditions. We applied QR codes as markers on the source videos to create aligned references and compute temporal features based on the alignment vectors. Using these features together with VMAF core features, our proposed model achieves a PCC of 0.99 on the validation set. Furthermore, our model outputs per-frame quality that gives detailed insight into the cause of video quality impairments. The VCM model and dataset are open-sourced at https://github.com/microsoft/Video_Call_MOS.

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.

Human subjective evaluation is the "gold standard" to evaluate speech quality optimized for human perception. Perceptual objective metrics serve as a proxy for subjective scores. We have recently developed a non-intrusive speech quality metric called Deep Noise Suppression Mean Opinion Score (DNSMOS) using the scores from ITU-T Rec. P.808 [1] subjective evaluation. The P.808 scores reflect the overall quality of the audio clip. ITU-T Rec. P.835 [2] subjective evaluation framework gives the standalone quality scores of speech and background noise in addition to the overall quality. In this work, we train an objective metric based on P.835 human ratings that output 3 scores: i) speech quality (SIG), ii) background noise quality (BAK), and iii) the overall quality (OVRL) of the audio. The developed metric is highly correlated with human ratings, with a Pearson’s Correlation Co-efficient (PCC)=0.94 for SIG and PCC=0.98 for BAK and OVRL. This is the first non-intrusive P.835 predictor we are aware of. DNSMOS P.835 is made publicly available as an Azure service.

With the recent growth of remote work, online meetings often encounter challenging audio contexts such as background noise, music, and echo.Accurate real-time detection of music events can help to improve the user experience.In this paper, we present MusicNet, a compact neural model for detecting background music in the real-time communications pipeline.In video meetings, music frequently co-occurs with speech and background noises, making the accurate classification quite challenging.We propose a compact convolutional neural network core preceded by an in-model featurization layer.MusicNet takes 9 seconds of raw audio as input and does not require any model-specific featurization in the product stack.We train our model on the balanced subset of the Audio Set [1] data and validate it on 1000 crowd-sourced real test clips.Finally, we compare MusicNet performance with 20 state-of-the-art models.MusicNet has a true positive rate (TPR) of 81.3% at a 0.1% false positive rate (FPR), which is significantly better than state-of-the-art models included in our study.MusicNet is also 10x smaller and has 4x faster inference than the best performing models we benchmarked.

Deep learning approaches-while remarkably successful in audio enhancement-result in slower inference times which are prohibitive in real-time applications. We develop a practical strategy for compressing U-Net style deep neural network architectures. On deep noise suppression (DNS) models we achieve a state-of-the-art 7.25× inference speed up over the baseline CRUSE model, with a smooth model performance degradation. Our method is friendly to practitioners as it only requires setting a single compression parameter, while achieving non-uniform compression rates across layers. We report inference speed because a parameter or memory reduction does not necessitate speedup, and we measure model quality using an accurate non-intrusive objective speech quality metric.

The Deep Noise Suppression (DNS) challenge is designed to foster innovation in the area of noise suppression to achieve superior perceptual speech quality. We recently organized a DNS challenge special session at INTERSPEECH 2020 where we open-sourced training and test datasets for researchers to train their noise suppression models. We also open-sourced a subjective evaluation framework and used the tool to evaluate and select the final winners. Many researchers from academia and industry made significant contributions to push the field forward. We also learned that as a research community, we still have a long way to go in achieving excellent speech quality in challenging noisy real-time conditions. In this challenge, we expanded both our training and test datasets. Clean speech in the training set has increased by 200% with the addition of singing voice, emotion data, and non-English languages. The test set has increased by 100% with the addition of singing, emotional, non-English (tonal and non-tonal) languages, and, personalized DNS test clips. There are two tracks with focus on (i) real-time denoising, and (ii) real-time personalized DNS. We present the challenge results at the end.

Human subjective evaluation is the "gold standard" to evaluate speech quality optimized for human perception. Perceptual objective metrics serve as a proxy for subjective scores. The conventional and widely used metrics require a reference clean speech signal, which is unavailable in real recordings. Previous no-reference approaches correlate poorly with human ratings and are not widely adopted in the research community. One of the biggest use cases of these perceptual objective metrics is to evaluate noise suppression algorithms. This paper introduces a multi-stage self-teaching based perceptual objective metric that is designed to evaluate noise suppressors. The proposed method generalizes well in challenging test conditions with a high correlation to human ratings.

The Deep Noise Suppression (DNS) challenge is designed to foster innovation in the area of noise suppression to achieve superior perceptual speech quality. We recently organized a DNS challenge special session at INTERSPEECH and ICASSP 2020. We open-sourced training and test datasets for the wideband scenario. We also open-sourced a subjective evaluation framework based on ITU-T standard P.808, which was also used to evaluate participants of the challenge. Many researchers from academia and industry made significant contributions to push the field forward, yet even the best noise suppressor was far from achieving superior speech quality in challenging scenarios. In this version of the challenge organized at INTERSPEECH 2021, we are expanding both our training and test datasets to accommodate full band scenarios. The two tracks in this challenge will focus on real-time denoising for (i) wide band, and(ii) full band scenarios. We are also making available a reliable non-intrusive objective speech quality metric called DNSMOS for the participants to use during their development phase.

The INTERSPEECH 2020 Deep Noise Suppression (DNS) Challenge is intended to promote collaborative research in real-time single-channel Speech Enhancement aimed to maximize the subjective (perceptual) quality of the enhanced speech. A typical approach to evaluate the noise suppression methods is to use objective metrics on the test set obtained by splitting the original dataset. While the performance is good on the synthetic test set, often the model performance degrades significantly on real recordings. Also, most of the conventional objective metrics do not correlate well with subjective tests and lab subjective tests are not scalable for a large test set. In this challenge, we open-sourced a large clean speech and noise corpus for training the noise suppression models and a representative test set to real-world scenarios consisting of both synthetic and real recordings. We also open-sourced an online subjective test framework based on ITU-T P.808 for researchers to reliably test their developments. We evaluated the results using P.808 on a blind test set. The results and the key learnings from the challenge are discussed. The datasets and scripts can be found here for quick access https://github.com/microsoft/DNS-Challenge.

Web-scale applications can ship code on a daily to weekly cadence. These applications rely on online metrics to monitor the health of new releases. Regressions in metric values need to be detected and diagnosed as early as possible to reduce the disruption to users and product owners. Regressions in metrics can surface due to a variety of reasons: genuine product regressions, changes in user population and bias due to telemetry loss (or processing) are among the common causes. Diagnosing the cause of these metric regressions is costly for engineering teams as they need to invest time in finding the root cause of the issue as soon as possible. We presentLumos, a Python library built using the principles of A/B testing to systematically diagnose metric regressions to automate such analysis.Lumos has been deployed across the component teams in Microsoft's Real-Time Communication (RTC) applications Skype and Microsoft Teams. It has enabled engineering teams to detect 100s of real changes in metrics and reject 1000s of false alarms detected by anomaly detectors. The application ofLumos has resulted in freeing up as much as $95%$ of the time allocated to metric-based investigations. In this work, we open sourceLumos and present our results from applying it to two different components within the RTC group over millions of sessions. This general library can be coupled with any production system to manage the volume of alerting efficiently.