Acoustic Echo Cancellation (AEC): Why Call Quality Depends on It
Written by
Aahan Sawhney
Category
Contact Center Ops
Date
Share this article
What Acoustic Echo Cancellation Actually Does to Your Calls
Picture this: a healthcare patient calls to confirm an appointment. The AI voice agent answers instantly, begins its greeting, and within seconds the caller hears their own voice bouncing back at them with a half-second delay. They repeat themselves. The agent misinterprets the echo as a new utterance. The conversation collapses into a loop of confused repetition, and the caller hangs up. No appointment confirmed. No record updated. Just a failed interaction that erodes trust in every future call.
That failure had nothing to do with the AI model, the script, or the integration. It was a signal processing problem. Specifically, it was an acoustic echo cancellation failure.
Acoustic echo cancellation (AEC) is the real-time signal processing technique that removes a speaker's own audio output from the microphone signal before that signal is transmitted back to the remote party. In plain terms, it stops your phone from sending your own voice back to the person you are talking to. In AI voice agent deployments, it also stops the agent from hearing its own synthesized speech looped back through the caller's device, which would otherwise be misread as new caller input.
This distinction matters enormously in AI-driven conversations. Human-to-human calls can tolerate a small amount of echo because both parties instinctively adjust, slow down, or ask for clarification. An AI voice agent has no such social grace. It processes what arrives in the audio stream and acts on it. If the agent's own words come back 150 to 400 milliseconds later, the speech recognition layer may treat them as a new utterance, triggering an unintended response, an interruption, or a complete derailment of the conversation flow.
Why Echo Exists in the First Place
Echo in telephony has two primary sources. Acoustic echo originates in the physical environment. A caller's phone speaker plays audio, that audio bounces off walls or hard surfaces, and the device microphone picks it up along with the caller's voice. The louder the speaker output and the smaller or harder the room, the worse the echo. Line echo is electrical in origin, generated by impedance mismatches in the telephone network, particularly at the hybrid junction where a four-wire circuit meets a two-wire local loop. Both types produce the same perceptual problem for the listener: hearing their own voice or the remote party's voice reflected back at them.
Modern VoIP and cloud telephony have largely pushed line echo to the margins. But acoustic echo remains a very real problem, especially as callers move to speakerphone mode on mobile devices, use smart speakers, or participate in calls from open-plan environments. For AI voice agent platforms, acoustic echo carries an additional layer of risk because the agent is always "listening" while it is also speaking, a state that humans avoid naturally by holding the phone to their ear.
The Signal Processing Foundation
At its core, AEC works by maintaining a reference signal of what the loudspeaker is currently outputting, then using an adaptive filter to model the acoustic path between that speaker and the microphone, and finally subtracting the predicted echo from the microphone signal in real time. The result is called the error signal, which ideally contains only the near-end speaker's voice (or silence) with the echo removed.
The adaptive filter is the critical component. It must continuously update itself because the acoustic path changes constantly. A caller shifting their grip on the phone, moving from a hard-floored hallway to a carpeted room, or covering the microphone briefly will all change the echo characteristics. Algorithms such as Normalized Least Mean Squares (NLMS) and more modern deep learning variants are designed to track these changes in milliseconds without introducing audible artifacts into the cleaned signal.
Where AEC gets genuinely difficult is in double-talk scenarios, where both parties are speaking simultaneously. The adaptive filter must detect that new near-end speech has arrived and temporarily halt or slow its adaptation, because if it keeps trying to model the acoustic path while both signals are active, it may begin canceling the caller's real voice instead of the echo. Double-talk detection is a sub-problem within AEC that separates basic implementations from production-grade ones.
Why This Is Especially Critical for AI Voice Agents
Traditional contact center software was designed around human agents wearing headsets. A headset places a physical barrier between speaker and microphone, dramatically reducing acoustic coupling and making AEC relatively straightforward. AI voice agents have no headset. The audio architecture is entirely software-defined, running on cloud infrastructure that may route through multiple network hops before reaching the caller's device.
In this environment, AEC must happen at the platform level, not the device level. Every millisecond of latency in the cancellation pipeline adds to the risk of the agent misprocessing its own reflected output. And because AI voice agents are expected to handle hundreds or thousands of concurrent calls in production deployments, the AEC layer must scale horizontally without degrading quality at volume.
For businesses in financial services, healthcare, and insurance, the stakes are compounded by compliance requirements. A call in which a patient's appointment details are garbled by echo, or in which a loan officer's instructions are echoed back and misprocessed, creates not only a poor experience but a potential documentation and liability problem. HIPAA-covered entities and financial institutions operating under consumer protection regulations need to know that every call is intelligible, every exchange is accurately captured, and every recording is clean enough to audit.
"Call quality is not a nice-to-have for regulated industries. It is a compliance requirement. If the recording is unintelligible, the interaction never happened for audit purposes."
This is why enterprise-grade voice AI platforms treat AEC as a foundational infrastructure requirement, not an optional add-on or a tuning parameter left to the deploying business to configure. When a business deploys an AI voice agent to handle inbound patient calls or outbound insurance follow-ups, they are implicitly relying on the platform to have solved echo cancellation at a level that survives real-world call conditions: speakerphone callers, noisy environments, network jitter, and concurrent call volume at scale.
The Business Cost of Getting It Wrong
Echo-related call quality failures are rarely logged as "echo problems" in a contact center's reporting. They surface as high abandonment rates, elevated repeat call volume, low caller satisfaction scores, and in AI deployments specifically, as unexpected transfer rates where callers escalate to a human because the automated interaction became incoherent.
For a business running 500 or 1,000 AI-handled calls per day, even a 3 to 5 percent failure rate attributable to poor echo cancellation represents dozens of lost interactions daily, each one a potential revenue or retention loss. In healthcare, a failed appointment confirmation call may mean a no-show slot and a patient who didn't receive timely care. In insurance, a failed outbound follow-up may mean a policy renewal lost to a competitor who got through clearly.
The businesses most exposed to this risk are those that have moved call volume to AI agents without auditing the underlying audio quality infrastructure. They may have evaluated the AI model's language understanding, the CRM integration, the call script logic, and the escalation rules, but if nobody tested what happens to audio quality when ten callers simultaneously use speakerphone in reverberant environments, the deployment has a hidden failure mode waiting to surface at scale.
Understanding what acoustic echo cancellation is and how it works is the first step toward asking the right questions of any voice AI platform before signing a contract or going live with production call volume.
How Acoustic Echo Cancellation Works Inside a Voice AI Stack
Most discussions of acoustic echo cancellation treat it as a single technology. In a production voice AI stack, it is more accurately a pipeline of interconnected processes, each of which must perform correctly for the output to be clean. Understanding this pipeline helps operations leaders ask better questions when evaluating voice AI platforms and diagnose what is actually going wrong when call quality degrades.
The Four-Stage AEC Pipeline
Stage 1: Reference Signal Capture
The first requirement is knowing exactly what the loudspeaker is outputting at any given moment. This reference signal is the ground truth against which the adaptive filter will model the echo. In a hardware phone, the reference is captured at the digital-to-analog converter before the signal reaches the speaker. In a software-based AI voice agent platform, the reference is the synthesized speech waveform generated by the text-to-speech engine, which is a cleaner and more controllable signal than hardware-captured audio.
This is one genuine advantage of AI voice agent platforms over legacy hardware endpoints: the platform has perfect knowledge of what it just played because it generated the audio itself. There is no guesswork about what the speaker output. This makes the adaptive filter's job substantially easier in controlled conditions, though it does not eliminate the problem of acoustic reflection and variable room acoustics on the caller's end.
Stage 2: Adaptive Filter and Echo Path Modeling
The adaptive filter continuously estimates the acoustic transfer function between the speaker and the microphone. In mathematical terms, it models the impulse response of the room, which describes how a sound played through the speaker will arrive at the microphone after reflections, absorption, and delay. This model is updated on every processing frame, typically at intervals of 10 to 20 milliseconds, using algorithms that minimize the residual echo in the error signal.
The most widely deployed classical algorithm is Normalized Least Mean Squares (NLMS), which balances computational efficiency against convergence speed. More recent implementations use frequency-domain adaptive filters to handle long echo tails more efficiently, and an increasing number of commercial platforms now layer deep neural network (DNN) postfilters on top of classical AEC to suppress residual echo that the adaptive filter leaves behind. The DNN approach can learn non-linear echo characteristics that classical filters struggle with, such as distortion introduced by low-quality speakers or microphone clipping.
Stage 3: Double-Talk Detection
Double-talk, where the near-end speaker (the caller) is talking at the same time as the far-end speaker (the AI agent), is the hardest scenario for AEC. The adaptive filter needs the microphone signal to contain only echo (or silence) when it is updating its model. If the caller is also speaking, the filter may interpret the caller's voice as error signal and begin canceling it, a catastrophic outcome known as near-end speech cancellation.
Production-grade double-talk detectors use statistical methods to continuously estimate the probability that near-end speech is present. When detected, the filter's adaptation rate is reduced or frozen. The challenge is that this detection must happen within a few milliseconds, or the damage is already done. In AI voice agent contexts, double-talk is frequent because the agent may not pause speech in time when a caller interrupts, making this one of the most practically important quality indicators to test before deployment.
Stage 4: Residual Echo Suppression and Postfiltering
No adaptive filter achieves perfect cancellation. Some residual echo always remains, particularly during rapid changes in the acoustic environment. A residual echo suppressor applies spectral subtraction or Wiener filtering to further attenuate the residual signal, at the cost of some background noise coloring. The goal is to push residual echo below the perceptual threshold where callers can detect it.
In DNN-enhanced pipelines, a learned postfilter operates on the output of the classical suppressor and applies a time-frequency mask to selectively attenuate frequency bands that likely contain residual echo. This approach can achieve perceptual quality substantially above what classical methods alone can reach, which is why it is now standard in enterprise-grade voice AI infrastructure.
How Leading Voice AI Platforms Handle AEC
The way different platforms handle AEC reflects their broader architectural philosophy.
Vapi is a developer-first platform that gives engineering teams direct control over the audio processing stack, including choice of transport protocol and configuration of audio processing parameters. For teams with strong DSP expertise, this flexibility is valuable. For businesses that do not have signal processing engineers on staff, it means AEC configuration is another task that falls to the deploying team rather than being handled at the platform level. Vapi's audience tends to be technical builders who want to assemble their own stack, and audio processing is one more layer they get to own and tune.
Retell AI sits at a middle point between developer tool and business platform. It offers more managed infrastructure than Vapi but still requires meaningful technical involvement to configure and deploy. AEC behavior in Retell deployments is somewhat dependent on the underlying telephony provider chosen by the deploying team, which means quality can vary based on how the stack is assembled.
Bland AI is optimized for high-volume outbound calling at speed. Its AEC implementation is designed for the specific acoustic conditions of automated outbound calls, where the near-end is typically a consumer mobile device and double-talk is relatively infrequent. This optimization works well in its target use case but may be less thoroughly tuned for the inbound scenarios common in healthcare and insurance, where callers are in variable environments and conversations are less scripted.
Feather AI treats audio quality, including echo cancellation, as a platform-level responsibility rather than a deployment-level configuration task. Businesses deploying Feather AI agents for inbound patient scheduling, financial services inquiries, or insurance follow-ups do not need to configure AEC parameters or choose between audio processing libraries. The platform handles this at the infrastructure layer so that the quality delivered in a controlled test environment is consistent with what callers experience at production volume. This approach is specifically designed for operations leaders who are accountable for call quality outcomes without necessarily having signal processing expertise in their teams.
The Network Layer Interaction
AEC does not operate in isolation. Its performance is directly affected by network conditions because variable packet delay (jitter) and packet loss change the timing relationship between the reference signal and the microphone signal. If packets arrive out of order, the adaptive filter may apply corrections to the wrong part of the echo tail, leaving audible artifacts.
Production voice AI platforms manage this through jitter buffers that smooth out arrival timing before the AEC pipeline sees the signal, and through packet loss concealment (PLC) algorithms that reconstruct missing audio segments. The interaction between jitter buffer depth, PLC quality, and AEC convergence speed is a key determinant of real-world call quality under adverse network conditions, which is why testing in simulated poor-network environments is essential before a production launch.
Testing AEC Before Going Live
The gap between demo-environment call quality and production call quality is one of the most consistent failure points in voice AI deployments. In a demo, both parties are typically using headsets in quiet rooms on stable networks. In production, callers are on speakerphone in cars, kitchens, and open offices, on mobile networks with variable latency, and sometimes with background noise levels that challenge even the most sophisticated AEC implementations.
Platforms that offer pre-production testing against simulated caller personas allow teams to evaluate audio quality under realistic conditions before the first real caller is affected. This kind of testing should include at minimum: speakerphone scenarios with reverberant room simulation, double-talk stress testing at various overlap durations, and network degradation scenarios with simulated jitter and packet loss. Any platform that does not surface these testing capabilities before go-live is asking you to use production callers as your QA environment.
For regulated industries where every call may be subject to audit, this is not an acceptable risk.
Where Acoustic Echo Cancellation Falls Short and When Manual Processes Win
Honest evaluation of acoustic echo cancellation requires acknowledging that it is not a solved problem. Despite decades of research and significant advances from deep learning postfiltering, there are real-world conditions under which even the best AEC implementations produce audible artifacts, miss residual echo, or actively degrade call quality. Operations leaders deploying voice AI at scale need to know these limits before they discover them through caller complaints.
The Honest Limits of AEC Technology
Extreme Reverb Environments
AEC adaptive filters model the acoustic path using a finite impulse response filter of fixed or adaptive length. In highly reverberant environments, such as a caller in a tiled bathroom, an empty conference room with glass walls, or a large open warehouse, the echo tail can extend to 500 milliseconds or longer. Many AEC implementations are tuned for tail lengths of 100 to 200 milliseconds, which covers typical home and office environments. When a caller's environment produces a longer tail, the filter's model is simply too short to capture all the echo energy, and residual echo artifacts appear even after suppression.
This is not a failure of any specific platform. It is a physical constraint of the signal processing architecture. No amount of tuning will eliminate all echo from a caller who is using a speakerphone in a bathroom with 800 milliseconds of reverb tail. The practical mitigation is residual echo suppression, which can reduce the artifact to below perceptual threshold in mild cases, but in extreme cases the caller still hears degraded audio.
Nonlinear Distortion from Speaker Hardware
AEC adaptive filters assume that the acoustic path between speaker and microphone is linear, meaning the echo at the microphone is a filtered and delayed version of the speaker output. In practice, consumer device speakers, particularly low-cost smartphone speakers at high volume, introduce nonlinear distortion. The echo at the microphone contains harmonic components that were not in the reference signal, and the linear adaptive filter cannot model or subtract them. This leaves behind a residual that sounds like a buzzy or distorted version of the agent's speech.
DNN postfilters can partially address this because they learn from data and can model non-linear residuals, but they are not perfect. Callers using older or lower-quality devices in speakerphone mode are the most exposed to this artifact.
Rapid Acoustic Environment Changes
The adaptive filter takes a finite amount of time to converge on a new acoustic model when conditions change. If a caller picks up their phone (shifting from speakerphone to earpiece), moves to a different room, or walks past a hard wall, the acoustic path changes faster than the filter can track. During this reconvergence period, which typically lasts 50 to 200 milliseconds depending on the algorithm, there is a window of elevated echo. For most calls, this is barely perceptible. For AI voice agents that are continuously parsing the audio stream, it can produce misrecognition events at exactly the moments of physical transition.
Double-Talk Detection Lag
As described in the previous section, double-talk detection is critical. But even the best detectors have a response latency. When a caller begins speaking at the same moment as the AI agent, the first 10 to 30 milliseconds of that overlap may not trigger the double-talk detector in time, allowing a brief period of filter adaptation on a contaminated signal. In practice, this produces occasional clipping or attenuation of the first syllable of a caller's interruption. For callers who interrupt frequently (which is common in frustrated or urgent callers), this can make the agent seem unresponsive or mis-hearing.
Network-Induced Reference Signal Mismatch
In cloud-based voice AI platforms, the synthesized speech (the reference signal) is generated in one data center and the microphone signal arrives from the PSTN or VoIP network through another path. Even small timing mismatches between these two paths, measured in tens of milliseconds, reduce the effectiveness of the adaptive filter because it is subtracting an echo prediction that is slightly offset in time from the actual echo. Robust implementations use synchronization algorithms to align the reference and microphone signals before adaptive filtering, but this adds latency, and getting the alignment right under variable network conditions is genuinely hard.
Where Human Agents Still Have a Quality Advantage
This section needs to be direct: for specific call scenarios, human agents with proper headsets in quiet call center environments still deliver better audio quality than AI voice agent platforms with software-defined AEC. The physics of a close-talk microphone in a noise-canceling headset are simply favorable in ways that software cannot fully replicate when the caller is the source of acoustic challenge.
Specifically, high-stakes, emotionally sensitive calls in healthcare settings, such as delivering diagnosis information, discussing end-of-life care options, or managing an extremely distressed patient, benefit from the audio clarity that a human agent in a controlled environment provides. The margin for misunderstanding in these conversations is essentially zero, and even mild echo artifacts can contribute to miscommunication at moments of high emotional intensity. This is not a criticism of voice AI technology. It is an honest statement about where the human-plus-controlled-environment combination outperforms software AEC in the most demanding acoustic conditions.
Similarly, calls that involve reading back sensitive numbers, such as account balances, social security number confirmations, or prescription dosages, are particularly sensitive to echo artifacts because a single digit misheard due to echo contamination can have serious consequences. These call types warrant careful consideration of whether a human agent or a carefully tested AI agent with verified audio quality is the appropriate choice.
Common Deployment Mistakes That Make AEC Worse
Not testing with representative devices. Most voice AI platforms are tested on headsets during development. Production callers use smartphones, smart speakers, and low-cost cordless phones. A deployment that has never been tested with speakerphone simulation is accepting unknown risk.
Assuming demo quality equals production quality. Demo environments are acoustically controlled. Production is not. Teams should insist on pre-production testing in degraded acoustic conditions before signing off on a deployment.
Ignoring volume-level acoustic feedback. Some businesses configure their AI agents with synthesized voices at high output volume to sound "clear and confident." High speaker volume increases the acoustic coupling between speaker and microphone on the caller's device, making AEC harder. Calibrating the output volume to the minimum needed for intelligibility actually improves AEC performance.
Skipping call quality monitoring post-launch. AEC performance degrades over time as platform updates, telephony provider changes, and caller demographic shifts alter the acoustic conditions the filter was tuned for. Continuous call quality monitoring is not optional in production AI deployments.
Conflating echo with latency. Operations teams sometimes report "echo" when the actual problem is high latency causing callers to hear their own voice looped back through the network rather than through acoustic reflection. These are different problems with different solutions. Misdiagnosing one as the other leads to tuning the wrong component.
"The most dangerous assumption in voice AI deployments is that call quality in the demo translates directly to call quality at production scale. It rarely does without deliberate testing and monitoring."
Acknowledging these limits is not a reason to avoid deploying AI voice agents. It is a reason to choose platforms and deployment processes that take audio quality seriously as an engineering problem, not just a marketing checkbox.
How Feather AI Addresses Echo Cancellation for Regulated Industries
For operations leaders in financial services, healthcare, and insurance, deploying AI voice agents is not a purely technological decision. It is an operational and compliance decision. The questions are not just "can the AI understand callers?" but "can it understand callers reliably, across all the devices and environments our specific customer base uses, at the volume we need, in a way that holds up to audit?" Acoustic echo cancellation sits directly at the center of all three of those questions.
How Feather AI Fits This Problem
Feather AI is built as a production-ready AI voice agent platform for businesses in regulated industries. Its approach to audio quality, including echo cancellation, reflects that positioning in three specific ways.
Platform-Level Audio Quality Management
Feather AI handles echo cancellation and audio quality at the platform infrastructure layer, not as a configuration task delegated to the deploying business. When a healthcare organization or insurance carrier deploys Feather AI agents to handle inbound calls, they do not need to select an AEC library, configure filter parameters, or tune postfilter aggressiveness. The platform delivers consistent, production-grade audio quality as part of the standard offering. This matters practically because the operations leaders deploying these agents, heads of revenue operations, patient experience directors, call center managers, are accountable for call quality outcomes but typically do not have signal processing engineers on their teams.
Pre-Production Testing Against Simulated Caller Personas
One of Feather AI's confirmed capabilities is pre-production testing against simulated caller personas, which allows teams to stress-test agent behavior, including audio quality, before any real callers are affected. For echo cancellation specifically, this means teams can evaluate how their deployed agent performs when callers simulate speakerphone use, variable network conditions, and high double-talk scenarios before the first live call. This is the testing gap that causes the most post-launch quality surprises, and having it built into the deployment workflow is directly relevant to call quality assurance in regulated industries.
Real-Time Observability and Call Quality Monitoring
Feather AI provides real-time observability and call quality monitoring as part of the platform. This gives operations teams the ability to detect audio quality degradation, including echo-related artifacts, at the call level and in aggregate across high-volume deployments. For businesses running hundreds or thousands of AI-handled calls per day, the ability to see quality metrics in real time and identify patterns (for example, elevated echo complaints correlated with a specific telephony carrier route or time of day) is essential to maintaining the kind of call quality that regulated industry compliance demands.
These three capabilities work together. Platform-level AEC means the baseline is sound. Pre-production testing means the baseline is verified under realistic conditions before go-live. Real-time monitoring means quality regressions are caught and addressed before they accumulate into a compliance or satisfaction problem.
The Nada Deployment as a Signal Quality Reference
Feather AI's deployment with Nada, a real estate investment platform, provides a relevant proof point on call quality at volume. Nada was receiving 40 or more inbound leads per day that were going cold because the sales team could not respond fast enough. Feather AI deployed an agent named "Jessica" for instant outreach, qualification, and warm transfer of qualified leads. In the first 30 days, the agent handled more than 5,000 calls and achieved a 19.5% warm transfer rate, meaning nearly one in five calls resulted in a successful handoff to a human sales rep with full context attached.
A warm transfer rate of 19.5 percent at 5,000-call volume is only achievable if the audio quality across those calls was consistently high enough for the AI to accurately qualify callers and for callers to trust the interaction enough to stay engaged. Echo artifacts that confuse the AI or frustrate callers produce abandonment, not warm transfers. While this deployment was in real estate rather than a regulated vertical, the call volume and quality requirements are directly relevant to what financial services and insurance firms can expect from the platform.
Sundance Brennan, Head of Revenue at Nada, confirmed the agent was live in under two weeks and delivered immediate, measurable impact on lead conversion.
Who Feather AI Is Not the Right Fit For
Being direct about fit is part of responsible evaluation. Feather AI is not the right choice for:
Solo developers or technical teams who want to assemble a fully custom voice stack and own every layer of audio processing configuration. That use case is better served by Vapi, which gives engineering teams direct access to the underlying audio infrastructure.
Very low-volume or non-regulated small businesses where the investment in production-grade audio quality infrastructure and the accompanying compliance posture (HIPAA, GDPR, SOC 2) is disproportionate to the call volume and risk profile.
Buyers expecting instant self-serve signup with no sales conversation. Feather AI is deployed through a structured onboarding process, not a credit-card-and-go product experience. Teams that need to be live tomorrow without any conversation will find this model frustrating.
For operations and revenue leaders at regulated businesses with real call volume (hundreds to thousands of calls per month) who need a working, compliant calling operation without hiring a team of signal processing and platform engineers, Feather AI is purpose-built for exactly that need.
Putting It Together: AEC as a Proxy for Platform Maturity
Acoustic echo cancellation is, in a meaningful sense, a proxy for the overall maturity of a voice AI platform. Getting AEC right requires solving a set of interconnected hard problems: adaptive signal processing, low-latency network architecture, real-world testing methodology, and ongoing quality monitoring. Platforms that have invested in solving these problems tend to have also invested in the other hard infrastructure problems that determine whether a voice AI deployment holds up at production scale in a regulated industry.
When evaluating voice AI platforms, asking specifically about how echo cancellation is handled, who configures it, how it is tested before go-live, and how quality is monitored post-launch, gives you a revealing window into how the vendor thinks about production readiness overall. A vendor that treats AEC as a solved checkbox is likely treating other hard production problems the same way. A vendor that can speak specifically about their approach to pre-production audio testing and real-time quality monitoring is demonstrating the operational discipline that regulated industry deployments actually require.
The calls your AI agents handle are not demos. They are real interactions with real patients, policyholders, and borrowers. The audio quality they experience is part of your brand, your compliance posture, and your conversion rate. Acoustic echo cancellation is the foundation that makes all of it possible.
Ready to see how Feather AI handles call quality in a production-grade deployment?
Share Blog
![](data:image/svg+xml,<svg display="block" role="presentation" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M 16 7.999 L 16 0 L 0 0 L 0 7.999 Z M 16 16 C 13.878 15.999 11.843 16.842 10.343 18.342 C 8.842 19.842 7.999 21.877 7.999 23.999 L 0.001 23.999 L 0.001 23.794 C 0.11 15.086 7.176 8.054 15.896 8 L 24 8 L 24 24 L 16 24 Z" fill="rgba(255, 255, 255, 0.03)" height="24px" id="lngUwvl9j" width="24.0000007312145px"/></svg>)