How Amazon Chime's noise cancellation works

Combining classic signal processing with deep learning makes method efficient enough to run on a phone.

PercepNet is one of the core technologies of Amazon Chime's Voice Focus feature. It is designed to suppress noise and reverberation in the speech signal, in real time, without using too many CPU cycles. This makes it usable in cellphones and other power-constrained devices. 

At Interspeech 2020, PercepNet finished second in its category (real-time processing) in the Deep Noise Suppression Challenge, despite using only 4% of a CPU core, while another Amazon Chime algorithm, PoCoNet, finished first in the offline-processing category. In this post, we'll look into the principles that make PercepNet work. For more details, you can also refer to our Interspeech paper.

Despite operating in real time, with low complexity, PercepNet can still provide state-of-the-art speech enhancement. Like most recent speech enhancement algorithms, PercepNet uses deep learning, but it applies it in a different way. Rather than have a deep neural network (DNN) do all the work, PercepNet tries to have it do as little work as possible.

Speech enhancement and STFT

Before getting into any deep learning, let's look at the job we'll be asking our machine learning model to perform. Let's consider a simple synthetic example. We start from the clean speech sample below:

We then add some non-stationary car noise on top of it:

The goal here is to take the noisy audio and make it sound as good as possible — ideally, close to the original clean audio. The standard way to represent the problem — both pre-deep learning and post-deep learning — is to use the short-time Fourier transform (STFT).

That means chopping up the signal into overlapping windows and computing the frequency content for each window. For each window of N samples (N discrete measurements of the signal amplitude), we obtain N/2 spectral magnitudes, along with their associated phases. We will refer to each output point as a frequency bin. Let's see what the magnitude of the STFT looks like for our clean signal (top) and noisy signal (bottom).

percepnet_spectrograms.jpg
The spectrograms above show the frequency content of an audio clip. The horizontal axis is time, the vertical axis is frequency, and the color represents the amount of energy at a particular time, for a particular frequency, using a log scale.

From the noisy STFT, many algorithms try to estimate the clean magnitude of each frequency while retaining the phase — which is much harder to estimate — from the noisy signal. For now, let's assume we have a magic model (an oracle) that's able to do a perfect mapping from noisy spectral magnitudes to clean. This is why we started from a synthetic example, so we can compute the oracle output. Based on oracle magnitudes but using the noisy phase, we can reconstruct the speech signal:

Certainly not bad, but also far from perfect. The noise is still audible as a form of roughness in the speech. This is due to the error in the phase, which we took from the noisy signal. While the ear is essentially insensitive to the absolute phase, what we perceive here is the inconsistency of the phase across frames. In other words, the way in which the phase changes over time still does matter.

Another issue for real-time, power-constrained operation is the number of frequency bins whose amplitudes we need to estimate. Assuming we use 20-millisecond windows, the STFT bins will be spaced 50 Hz apart. If we want to enhance all frequencies up to 20 kHz (the upper limit of human hearing), then our neural network will have to estimate 400 amplitudes, which is very computationally expensive.

Where do we go from here? If we want to improve quality, then we could also estimate phase. This is the no-compromise route taken by PoCoNet, which can get around the added complexity because it’s optimized to run on a GPU. For real-time applications on power-constrained devices, however, we can't realistically expect to have a very good phase estimator.

A perceptually relevant representation

If we want good speech quality, and we want our algorithm to run in real time on a CPU without instantly draining the battery, then we need to find a way to simplify the problem. We can do that by making the following assumptions:

  1. the general shape of the speech spectrum (a.k.a. the spectral envelope) is smooth; and 
  2. we perceive it with a nonlinear frequency resolution, corresponding to the human ear’s auditory filters (a.k.a. critical bands)

In other words, (1) the speech spectrum tends not to have sharp discontinuities, and (2) the human auditory system perceives low frequencies with higher resolution than high frequencies.

We can follow both of those assumptions by representing the speech spectrum using bands spaced according to equivalent rectangular bandwidth (ERB). ERB-spaced bands divide the spectrum into bands of increasing width, capturing coarser spectral information as frequency increases, much the way the human auditory system does.

Because multiple STFT bins are assigned to each band, the spectral representation is smoother: any discontinuity in frequency is averaged out.

Nonlinearly spaced bands make our model much simpler. Instead of 400 frequency bins, we need only 34 bands. In practice, we model these bands as overlapping filters, which are most responsive to the frequencies at the centers of the bands (the tips of the triangles below) and decreasingly responsive to frequencies farther from the center (the sides of the triangles; note the 50% overlap between bands):

bands.png

For each of the bands above, we compute a gain between 0 and 1; then, all we need to do is interpolate those band gains and we're done. Now, let's listen to how this would sound — still using the oracle for band magnitudes:

Our complexity went down, but so did the quality. The roughness we noticed previously is now even more obvious and sounds a bit like heavy distortion. It's not that surprising, since we are still changing only the magnitude spectrum, but with only 34 degrees of freedom rather than 400.

So what are we missing here? The missing piece is that the ear doesn't only perceive the spectral envelope of the signal; it also perceives whether the signal is made of tones (voiced sounds), noise (unvoiced sounds), or a mix of the two. Vowels are mostly composed of tones (harmonics) at multiples of a fundamental frequency (the pitch), whereas many consonants (such as the /s/ phoneme) are mostly noise-like. 

Our enhanced speech sounds rough because the tonal vowels contain more noise than they should. To enhance our tones, we can use a time-domain technique called comb filtering. Comb filtering is often an undesired effect in which room reverberation boosts or attenuates frequencies at regular intervals. But by carefully tuning our comb filter to the pitch of the voice we're trying to enhance, we can keep all the tones and remove most of the noise. Below is an example of the frequency response of the comb filter for a pitch of 200 Hz.

pitch.png

The pitch is the period at which a periodic signal (nearly) repeats itself. Pitch estimation is a hard problem, especially in the noisy conditions we have here. To estimate the pitch, we try to match a signal with past versions of itself, finding the period T that maximizes the correlation between x(n) and x(n-T). We then use dynamic programming (the Viterbi algorithm) to find a pitch trajectory that is consistent (e.g. no large jumps) over time.

Since we often want to retain at least some of the noise, we can simply do a mix between the noisy audio and the comb-filtered audio to get exactly the tone/noise ratio we want. By doing the mixing in the frequency domain, we can control that mix on a band-by-band basis, even though the comb filter is computed in the time domain. The exact ratios (or filtering strengths) to use for the mixing can be adjusted in such a way that the ratio of tones to noise in the output is about the same as it was in the clean speech. This is what our oracle (using the optimal strengths) now sounds like with comb filtering:

There’s still a little roughness, but our quality is already better than that of our spectral-magnitude oracle, despite using far fewer parameters. It now seems that we're as close to the original properties of the speech as we could get with our model. So what else can we do to further improve quality? The answer is simple: we cheat! 

To be more specific, we can cheat the human auditory system a bit by further attenuating the frequency bands that are still too noisy. Our speech will deviate slightly from the correct spectral envelope, but the ear will not notice that too much. It will just notice the noise less. This kind of post-filtering has been used in speech codecs since the 1980s but (as far as we know) not in speech enhancement systems. Adding the post-filter to our oracle gives us the following:

We're now quite close to the perfect clean speech. At this point, our limiting factor will most certainly be the DNN model and not the representation we use. The good thing is that our DNN has to estimate only 34 band gains (between 0 and 1) and 34 comb-filtering strengths (also between 0 and 1). This is much easier than estimating 400 magnitudes/gains — and possibly also 400 phases.

Adding a DNN

So far, we’ve assumed a perfect model for predicting band gains (the oracle). In practice, we need to use a DNN. But all the work we did in the previous section was meant to make the DNN design as boring as possible.

Since we replaced our initial 400 frequency bins with just 34 bands, there's no reason to use convolutional layers across frequency. Instead, we just go with convolutional layers across time and — most importantly — recurrent layers that provide longer-term memory to the system. We found that simple gated recurrent units (GRUs) work well, but long-short-term-memory networks (LSTMs) would probably have worked as well.

dnn_model.png
DNN model

In our DNN modelf is an input feature vector that contains all the band-based spectral information we need. The outputs are the band gains b and the comb-filtering strengths b. Now all we need to do is train our network using hours of clean speech to which we add various levels of noise and reverberation. Since we have the clean speech, we can compute the optimal (oracle) gains and filtering strengths and use them as training targets. Our complete system using the trained DNN sounds like this:

Obviously, it does not sound as good as the last oracle — no enhancement DNN is perfect — but it's still a big improvement over the noisy input speech. Our Interspeech 2020 Deep Noise Suppression Challenge samples page provides some examples of how PercepNet performs in real conditions.

Using it in real time

The DNN model above contains about eight million weights. For each new window, we use each weight exactly once, which means eight million multiply-add operations per window. With 20-millisecond windows and 50% overlap, we have 100 windows per second of speech, so 800 million multiply-add operations per second. 

Thankfully, DNNs tend to be quite robust to small perturbations, so we can quantize all our weights to just eight bits with a negligible effect on perceived audio quality. Thanks to SIMD instructions on modern CPUs, this makes it possible to run our network really efficiently. On a modern laptop CPU, it takes less than 5% of one core to run PercepNet in real time.

To be useful in real-time communications applications, PercepNet should not add too much delay. The seemingly arbitrary choice of 20-millisecond windows with 50% overlap means that it consumes audio 10 milliseconds at a time. This is good because most audio codecs (including Opus, which is used in WebRTC) encode audio in 20-millisecond packets. So we can run the algorithm exactly twice per packet without the PercepNet block size causing an increase in delay. 

There are, of course, some delays we cannot avoid. The overlap between windows means that the STFT itself requires 10 milliseconds for reconstruction. On top of that, we typically allow the DNN to look two windows (20 millseconds) into the future, so it can make better decisions. This gives us a total of 30 milliseconds extra delay from the algorithm, which is acceptable in most scenarios.

If you would like to know more about the details of PercepNet, you can read our Interspeech 2020 paper. The idea behind PercepNet is quite versatile and could be applied to other problems, including acoustic echo control and beamforming post-filtering. In future posts, we will see how we can make PercepNet very efficient on CPUs and even how to run it as Web Assembly (WASM) code inside web browsers for WebRTC-based applications.

Related content

US, CA, San Francisco
About Twitch Launched in 2011, Twitch is a global community that comes together each day to create multiplayer entertainment: unique, live, unpredictable experiences created by the interactions of millions. We bring the joy of co-op to everything, from casual gaming to world-class esports to anime marathons, music, and art streams. Twitch also hosts TwitchCon, where we bring everyone together to celebrate, learn, and grow their personal interests and passions. We’re always live at Twitch. Stay up to date on all things Twitch on Linkedin, Twitter and on our Blog. About the role: Twitch builds data-driven machine learning solutions across several rich problem spaces: Natural Language Processing (NLP), Recommendations, Semantic Search, Classification/Categorization, Anomaly Detection, Forecasting, Safety, and HCI/Social Computing/Computational Social Science. As an Intern, you will work with a dedicated Mentor and Manager on a project in one of these problem areas. You will also be supported by an Advisor and participate in cohort activities such as research teach backs and leadership talks. This position can also be located in San Francisco, CA or virtual. You Will: Solve large-scale data problems. Design solutions for Twitch's problem spaces Explore ML and data research
LU, Luxembourg
Are you a talented and inventive scientist with a strong passion about modern data technologies and interested to improve business processes, extracting value from the data? Would you like to be a part of an organization that is aiming to use self-learning technology to process data in order to support the management of the procurement function? The Global Procurement Technology, as a part of Global Procurement Operations, is seeking a skilled Data Scientist to help build its future data intelligence in business ecosystem, working with large distributed systems of data and providing Machine Learning (ML) and Predictive Modeling expertise. You will be a member of the Data Engineering and ML Team, joining a fast-growing global organization, with a great vision to transform the Procurement field, and become the role model in the market. This team plays a strategic role supporting the core Procurement business domains as well as it is the cornerstone of any transformation and innovation initiative. Our mission is to provide a high-quality data environment to facilitate process optimization and business digitalization, on a global scale. We are supporting business initiatives, including but not limited to, strategic supplier sourcing (e.g. contracting, negotiation, spend analysis, market research, etc.), order management, supplier performance, etc. We are seeking an individual who can thrive in a fast-paced work environment, be collaborative and share knowledge and experience with his colleagues. You are expected to deliver results, but at the same time have fun with your teammates and enjoy working in the company. In Amazon, you will find all the resources required to learn new skills, grow your career, and become a better professional. You will connect with world leaders in your field and you will be tackling Data Science challenges to ensure business continuity, by taking the right decisions for your customers. As a Data Scientist in the team, you will: -be the subject matter expert to support team strategies that will take Global Procurement Operations towards world-class predictive maintenance practices and processes, driving more effective procurement functions, e.g. supplier segmentation, negotiations, shipping supplies volume forecast, spend management, etc. -have strong analytical skills and excel in the design, creation, management, and enterprise use of large data sets, combining raw data from different sources -provide technical expertise to support the development of ML models to facilitate intelligent digital services, such as Contract Lifecycle Management (CLM) and Negotiations platform -cooperate closely with different groups of stakeholders, e.g. data/software engineers, product/program managers, analysts, senior leadership, etc. to evaluate business needs and objectives to set up the best data management environment -create and share with audiences of varying levels technical papers and presentations -deal with ambiguity, prioritizing needs, and delivering results in a dynamic environment Basic qualifications -Master’s Degree in Computer Science/Engineering, Informatics, Mathematics, or a related technical discipline -3+ years of industry experience in data engineering/science, business intelligence or related field -3+ years experience in algorithm design, engineering and implementation for very-large scale applications to solve real problems -Very good knowledge of data modeling and evaluation -Very good understanding of regression modeling, forecasting techniques, time series analysis, machine-learning concepts such as supervised and unsupervised learning, classification, random forest, etc. -SQL and query performance tuning skills Preferred qualifications -2+ years of proficiency in using R, Python, Scala, Java or any modern language for data processing and statistical analysis -Experience with various RDBMS, such as PostgreSQL, MS SQL Server, MySQL, etc. -Experience architecting Big Data and ML solutions with AWS products (Redshift, DynamoDB, Lambda, S3, EMR, SageMaker, Lex, Kendra, Forecast etc.) -Experience articulating business questions and using quantitative techniques to arrive at a solution using available data -Experience with agile/scrum methodologies and its benefits of managing projects efficiently and delivering results iteratively -Excellent written and verbal communication skills including data visualization, especially in regards to quantitative topics discussed with non-technical colleagues
US, CA, San Francisco
About Twitch Launched in 2011, Twitch is a global community that comes together each day to create multiplayer entertainment: unique, live, unpredictable experiences created by the interactions of millions. We bring the joy of co-op to everything, from casual gaming to world-class esports to anime marathons, music, and art streams. Twitch also hosts TwitchCon, where we bring everyone together to celebrate, learn, and grow their personal interests and passions. We’re always live at Twitch. Stay up to date on all things Twitch on Linkedin, Twitter and on our Blog. About the role: Twitch builds data-driven machine learning solutions across several rich problem spaces: Natural Language Processing (NLP), Recommendations, Semantic Search, Classification/Categorization, Anomaly Detection, Forecasting, Safety, and HCI/Social Computing/Computational Social Science. As an Intern, you will work with a dedicated Mentor and Manager on a project in one of these problem areas. You will also be supported by an Advisor and participate in cohort activities such as research teach backs and leadership talks. This position can also be located in San Francisco, CA or virtual. You Will: Solve large-scale data problems. Design solutions for Twitch's problem spaces Explore ML and data research
US, CA, San Francisco
About Twitch Launched in 2011, Twitch is a global community that comes together each day to create multiplayer entertainment: unique, live, unpredictable experiences created by the interactions of millions. We bring the joy of co-op to everything, from casual gaming to world-class esports to anime marathons, music, and art streams. Twitch also hosts TwitchCon, where we bring everyone together to celebrate, learn, and grow their personal interests and passions. We’re always live at Twitch. Stay up to date on all things Twitch on Linkedin, Twitter and on our Blog. About the role: Twitch builds data-driven machine learning solutions across several rich problem spaces: Natural Language Processing (NLP), Recommendations, Semantic Search, Classification/Categorization, Anomaly Detection, Forecasting, Safety, and HCI/Social Computing/Computational Social Science. As an Intern, you will work with a dedicated Mentor and Manager on a project in one of these problem areas. You will also be supported by an Advisor and participate in cohort activities such as research teach backs and leadership talks. This position can also be located in San Francisco, CA or virtual. You Will: Solve large-scale data problems. Design solutions for Twitch's problem spaces Explore ML and data research
US, CA, San Francisco
About Twitch Launched in 2011, Twitch is a global community that comes together each day to create multiplayer entertainment: unique, live, unpredictable experiences created by the interactions of millions. We bring the joy of co-op to everything, from casual gaming to world-class esports to anime marathons, music, and art streams. Twitch also hosts TwitchCon, where we bring everyone together to celebrate, learn, and grow their personal interests and passions. We’re always live at Twitch. Stay up to date on all things Twitch on Linkedin, Twitter and on our Blog. About the role: Twitch builds data-driven machine learning solutions across several rich problem spaces: Natural Language Processing (NLP), Recommendations, Semantic Search, Classification/Categorization, Anomaly Detection, Forecasting, Safety, and HCI/Social Computing/Computational Social Science. As an Intern, you will work with a dedicated Mentor and Manager on a project in one of these problem areas. You will also be supported by an Advisor and participate in cohort activities such as research teach backs and leadership talks. This position can also be located in San Francisco, CA or virtual. You Will: Solve large-scale data problems. Design solutions for Twitch's problem spaces Explore ML and data research
US, CA, San Francisco
About Twitch Launched in 2011, Twitch is a global community that comes together each day to create multiplayer entertainment: unique, live, unpredictable experiences created by the interactions of millions. We bring the joy of co-op to everything, from casual gaming to world-class esports to anime marathons, music, and art streams. Twitch also hosts TwitchCon, where we bring everyone together to celebrate, learn, and grow their personal interests and passions. We’re always live at Twitch. Stay up to date on all things Twitch on Linkedin, Twitter and on our Blog. About the role: Twitch builds data-driven machine learning solutions across several rich problem spaces: Natural Language Processing (NLP), Recommendations, Semantic Search, Classification/Categorization, Anomaly Detection, Forecasting, Safety, and HCI/Social Computing/Computational Social Science. As an Intern, you will work with a dedicated Mentor and Manager on a project in one of these problem areas. You will also be supported by an Advisor and participate in cohort activities such as research teach backs and leadership talks. This position can also be located in San Francisco, CA or virtual. You Will: Solve large-scale data problems. Design solutions for Twitch's problem spaces Explore ML and data research
US, WA, Seattle
Amazon is seeking an experienced, self-directed data scientist to support the research and analytical needs of Amazon Web Services' Sales teams. This is a unique opportunity to invent new ways of leveraging our large, complex data streams to automate sales efforts and to accelerate our customers' journey to the cloud. This is a high-visibility role with significant impact potential. You, as the right candidate, are adept at executing every stage of the machine learning development life cycle in a business setting; from initial requirements gathering to through final model deployment, including adoption measurement and improvement. You will be working with large volumes of structured and unstructured data spread across multiple databases and can design and implement data pipelines to clean and merge these data for research and modeling. Beyond mathematical understanding, you have a deep intuition for machine learning algorithms that allows you to translate business problems into the right machine learning, data science, and/or statistical solutions. You’re able to pick up and grasp new research and identify applications or extensions within the team. You’re talented at communicating your results clearly to business owners in concise, non-technical language. Key job responsibilities • Work with a team of analytics & insights leads, data scientists and engineers to define business problems. • Research, develop, and deliver machine learning & statistical solutions in close partnership with end users, other science and engineering teams, and business stakeholders. • Use AWS services like SageMaker to deploy scalable ML models in the cloud. • Examples of projects include modeling usage of AWS services to optimize sales planning, recommending sales plays based on historical patterns, and building a sales-facing alert system using anomaly detection.
US, WA, Seattle
We are a team of doers working passionately to apply cutting-edge advances in deep learning in the life sciences to solve real-world problems. As a Senior Applied Science Manager you will participate in developing exciting products for customers. Our team rewards curiosity while maintaining a laser-focus in bringing products to market. Competitive candidates are responsive, flexible, and able to succeed within an open, collaborative, entrepreneurial, startup-like environment. At the leading edge of both academic and applied research in this product area, you have the opportunity to work together with a diverse and talented team of scientists, engineers, and product managers and collaborate with others teams. Location is in Seattle, US Embrace Diversity Here at Amazon, we embrace our differences. We are committed to furthering our culture of inclusion. We have ten employee-led affinity groups, reaching 40,000 employees in over 190 chapters globally. We have innovative benefit offerings, and host annual and ongoing learning experiences, including our Conversations on Race and Ethnicity (CORE) and AmazeCon (gender diversity) conferences. Amazon’s culture of inclusion is reinforced within our 14 Leadership Principles, which remind team members to seek diverse perspectives, learn and be curious, and earn trust Balance Work and Life Our team puts a high value on work-life balance. It isn’t about how many hours you spend at home or at work; it’s about the flow you establish that brings energy to both parts of your life. We believe striking the right balance between your personal and professional life is critical to life-long happiness and fulfillment. We offer flexibility in working hours and encourage you to find your own balance between your work and personal lives Mentor & Grow Careers Our team is dedicated to supporting new members. We have a broad mix of experience levels and tenures, and we’re building an environment that celebrates knowledge sharing and mentorship. Our senior members enjoy one-on-one mentoring and thorough, but kind, code reviews. We care about your career growth and strive to assign projects based on what will help each team member develop into a better-rounded engineer and enable them to take on more complex tasks in the future. Key job responsibilities • Manage high performing engineering and science teams • Hire and develop top-performing engineers, scientists, and other managers • Develop and execute on project plans and delivery commitments • Work with business, data science, software engineer, biological, and product leaders to help define product requirements and with managers, scientists, and engineers to execute on them • Build and maintain world-class customer experience and operational excellence for your deliverables
US, Virtual
The Amazon Economics Team is hiring Interns in Economics. We are looking for detail-oriented, organized, and responsible individuals who are eager to learn how to work with large and complicated data sets. Some knowledge of econometrics, as well as basic familiarity with Stata, R, or Python is necessary. Experience with SQL, UNIX, Sawtooth, and Spark would be a plus. These are full-time positions at 40 hours per week, with compensation being awarded on an hourly basis. You will learn how to build data sets and perform applied econometric analysis at Internet speed collaborating with economists, data scientists and MBAʼs. These skills will translate well into writing applied chapters in your dissertation and provide you with work experience that may help you with placement. Roughly 85% of interns from previous cohorts have converted to full time economics employment at Amazon. If you are interested, please send your CV to our mailing list at econ-internship@amazon.com.
US, WA, Seattle
Amazon internships are full-time (40 hours/week) for 12 consecutive weeks with start dates in May - July 2023. Our internship program provides hands-on learning and building experiences for students who are interested in a career in hardware engineering. This role will be based in Seattle, and candidates must be willing to work in-person. Corporate Projects (CPT) is a team that sits within the broader Corporate Development organization at Amazon. We seek to bring net-new, strategic projects to life by working together with customers and evolving projects from ZERO-to-ONE. To do so, we deploy our resources towards proofs-of-concept (POCs) and pilot programs and develop them from high-level ideas (the ZERO) to tangible short-term results that provide validating signal and a path to scale (the ONE). We work with our customers to develop and create net-new opportunities by relentlessly scouring all of Amazon and finding new and innovative ways to strengthen and/or accelerate the Amazon Flywheel. CPT seeks an Applied Science intern to work with a diverse, cross-functional team to build new, innovative customer experiences. Within CPT, you will apply both traditional and novel scientific approaches to solve and scale problems and solutions. We are a team where science meets application. A successful candidate will be a self-starter comfortable with ambiguity, strong attention to detail, and the ability to work in a fast-paced, ever-changing environment. As an Applied Science Intern, you will own the design and development of end-to-end systems. You’ll have the opportunity to create technical roadmaps, and drive production level projects that will support Amazon Science. You will work closely with Amazon scientists, and other science interns to develop solutions and deploy them into production. The ideal scientist must have the ability to work with diverse groups of people and cross-functional teams to solve complex business problems.