Krisp, NVIDIA Maxine, And Dolby — Build Vs Buy For Real-Time Voice Enhancement

Published 2026-06-02 · 16 min read · By Nikolay Sapunov, CEO at Fora Soft

Why This Matters

If your product runs live calls — video conferencing, telemedicine, an AI voice agent, a sales tool, a virtual classroom — background noise is the fastest way to make it feel cheap, and a voice-enhancement engine is the fix. The trap is that the choice looks like a quality contest ("which model sounds best?") when it is really an operations and cost decision ("which model can I run where my audio actually is, at a price that survives scale, from a vendor that will still exist next year?"). This article is for the product manager, founder, or engineering lead who has to make that call and defend it, without being bluffed by a polished demo. By the end you will understand what Krisp, NVIDIA Maxine, and Dolby each really sell, why two of them shifted under the industry's feet in 2024–2026, how the money works for "build" versus each "buy", and a simple set of questions that lead you to the right answer for your product. For the deeper question of where in a WebRTC call a suppressor physically attaches, see our companion article on real-time noise suppression in WebRTC; this one is about which engine to commit to.

First, What "Voice Enhancement" Means Here

Before comparing vendors, anchor the thing they all sell. A voice-enhancement engine, in this article, is software that takes a stream of speech audio and removes what you do not want to hear — steady noise like a fan, sudden noise like a door, room echo, and in the best engines, other people's voices — while keeping the main speaker sounding natural. The most basic job is called noise suppression, also written noise cancellation: turning the non-speech sound down. A stronger job, which only some engines do, is background voice cancellation, usually shortened to BVC: removing nearby human voices that are not the main speaker, so a colleague talking across the room does not end up in your call or your transcript.

Two facts about these engines decide everything downstream. The first is that they must work on raw, uncompressed sound — the actual audio samples — because the model reads which frequencies are voice and which are noise. The second is that this work costs computing power, and where that computing happens — on the speaker's own laptop and phone, or on your servers — is the single biggest difference between the vendors below. Hold both thoughts.

The Build-Vs-Buy Spectrum, Not A Binary

People say "build versus buy" as if there were two boxes. In voice enhancement there is a spectrum, and the vendors sit at different points on it. At one end you build: you take a free, open model and operate every part of it yourself. At the other end you buy a fully managed engine that runs on the user's device and asks nothing of your servers. In between sits a hybrid where you license a vendor's model but run it on your own hardware. Knowing which point a vendor occupies tells you more than any quality score.

Figure 1. Build versus buy is a spectrum measured by how much you operate — from open models you run end to end, through Maxine's license-the-model-run-it-yourself middle, to fully managed engines like Krisp that run on the user's device.

The axis in Figure 1 is deliberately not "how good does it sound". Quality matters, but it is the second question. The first is operational: every step rightward removes work and hardware from your plate and replaces it with a licence fee. Where you want to land depends on your budget, your team, the devices your audio comes from, and how much you trust a vendor to stay on course. We will place each name on this spectrum and then turn it into a decision.

Build: The Free End, Where You Are The Maintainer

The leftmost option is to take an open-source model and run the whole thing yourself. Two models dominate here. RNNoise is a tiny, free noise suppressor that compiles down to a few hundred kilobytes and runs in any browser; it is good on steady noise and weaker on hard cases like overlapping voices. DeepFilterNet is a heavier open model that sounds noticeably better, at the cost of more size and more processing. Both are genuinely free to licence — RNNoise under the permissive BSD licence, DeepFilterNet under MIT and Apache terms — which is why "build" is tempting.

The word "free" hides the real bill. When you build, you compile the model, wire it into the audio path, manage the buffering that real-time audio demands, test it across every browser and phone you support, and own the result forever. If a new browser version breaks your audio plumbing, that is your bug to fix. If users complain that the model misses a kind of noise, there is no vendor to escalate to — improving it is your project. The licence is free; the engineering, the maintenance, and the quality ceiling are yours. For a browser-only product with ordinary noise and an engineering team that wants control, this is the right and cheapest answer. For a team that wants clean audio to be someone else's job, it is a trap dressed as a saving. The internal mechanics of these models — how they compare on speed, quality, and licence — are covered in our noise suppression models deep-dive.

Buy, Fully Managed: Krisp Runs On The User's Device

Krisp is the name most teams hear first, and it sits at the far "buy" end of the spectrum for a specific reason: its model runs on the speaker's own device, so you operate no servers for it at all. Krisp is a Voice AI company founded in 2017 and based in Berkeley, California. By the company's own statement, audio never leaves the device — the model processes sound locally and uploads nothing — which is a strong privacy property for telemedicine and any regulated product. Its noise technology runs across more than two hundred million devices and processes over seventy-five billion minutes of voice each month, and it powers noise cancellation inside Discord, RingCentral, and Zoho, with partner integrations into the calling platforms Twilio and Daily.

What you actually buy from Krisp is a software development kit — an SDK, a pre-built code library you drop into your app — for Windows, macOS, Linux, Android, iOS, and browsers, the last through a WebAssembly build. The SDK does more than basic noise removal. It includes background voice cancellation, the model that strips out other people's voices; voice isolation tuned for AI voice agents so a bot is not interrupted by background chatter; accent conversion; and per-frame statistics you can monitor. In a live call its standard model adds roughly twenty-five milliseconds of delay, with a lighter model around fifteen — small enough to stay inside a natural conversation's timing budget, the sub-100-millisecond latency budget every real-time feature has to respect.

The catch is commercial, not technical. Krisp's developer SDK is sold as an enterprise licence — typically an annual contract with an upfront commitment, quoted through sales rather than published per minute. A competitor, ai-coustics, makes this critique openly, arguing that Krisp's pricing suits established enterprise deployments but can be rigid for a small startup that wants to download, test, and scale gradually. Treat that as a competitor's framing, not gospel — but the shape is real: Krisp is a buy decision you make deliberately, with a contract, not a credit card. What you get for it is production-grade quality on every platform with zero server operations, which for many products is exactly worth the price.

Buy The Model, Run It Yourself: NVIDIA Maxine (Now AI For Media)

NVIDIA's offering sits in the middle of the spectrum, and in 2026 it carries a new name worth stating clearly: NVIDIA Maxine has been renamed NVIDIA AI for Media. Existing access to its SDKs and services continues uninterrupted; the technology is the same, the brand is new. We will write "Maxine (AI for Media)" so the older name you may have searched for still connects.

Maxine is different from Krisp in the most important way: it does not run on the user's device, it runs on NVIDIA graphics cards — which means your servers, your cloud GPUs, or NVIDIA-class workstations. Its Audio Effects component delivers real-time noise removal, room-echo removal, audio super-resolution, and acoustic echo cancellation, and a separate model called Studio Voice lifts ordinary microphone audio toward studio quality. These are accelerated by the Tensor Cores inside NVIDIA RTX-class GPUs, and the SDK requires such a GPU plus at least ten gigabytes of memory to run. That hardware requirement is the whole story: Maxine buys you arguably the widest and highest-ceiling set of audio effects available, but only if you can put NVIDIA GPUs where your audio is.

The commercial model follows the hardware. You can prototype against NVIDIA's hosted API catalogue for free and request a ninety-day evaluation licence, but production use runs through NVIDIA AI Enterprise — a software subscription priced at roughly four thousand five hundred dollars per GPU per year, or about one dollar per GPU-hour in the cloud, on top of the GPU hardware or rental itself. That makes Maxine the heaviest "buy" to operate: you are not just licensing a model, you are running a fleet. The payoff is a complete enhancement pipeline — audio plus video super-resolution, background effects, and more — on infrastructure you control, which suits broadcast, large-scale streaming, and AI agents that already live on GPU servers. For the wider economics of running AI on GPUs in a video product, see our cost-model article.

A Numeric Example: What "Run It Yourself" Costs At Scale

Maxine's cost only becomes real when you put numbers on it, so here is a worked example with the arithmetic shown. Suppose you run an AI voice-agent platform and need to clean one thousand inbound audio streams at the same time, server-side, because the audio arrives from phones and devices you do not control. Maxine's audio effects run on GPUs, so the question is how many GPUs that fleet needs.

As an illustration — you must measure this for your own workload, because NVIDIA does not publish a single streams-per-GPU figure — assume one GPU comfortably handles one hundred concurrent real-time audio streams. Then:

1,000 streams ÷ 100 streams per GPU = 10 GPUs needed

Now apply the licence cost, which is sourced rather than assumed. NVIDIA AI Enterprise runs about $4,500 per GPU per year:

10 GPUs × $4,500 per GPU-year = $45,000 per year in software licences alone

That is before the GPUs themselves. Renting them in the cloud at roughly one dollar per GPU-hour, running continuously, adds:

10 GPUs × $1 per GPU-hour × 24 hours × 365 days = $87,600 per year in GPU rental

So a thousand-stream Maxine deployment lands near $132,600 per year in licences plus cloud GPUs, and you still operate the fleet. Now contrast the shape of the bill with Krisp, where the model runs on each user's own device: your server cost for the audio cleaning is effectively zero, and you instead pay an annual SDK licence whose price does not climb with every concurrent stream the way GPU rental does. The lesson is not that one number beats the other — it is that the shape of the cost differs. Maxine's cost scales with concurrent server load; Krisp's scales with a contract; build scales with engineering hours. Choose the shape that matches how your product grows.

The Cautionary Tale: Dolby Moved, And A "Buy" Option Closed

Dolby belongs in any "Krisp versus Maxine versus Dolby" comparison because teams still search for it — and because what happened to it is the single most important build-versus-buy lesson on this page. For years, Dolby's developer platform, Dolby.io, offered self-service real-time Communications APIs built on technology called Voxeet, and those calls carried Dolby Voice noise suppression. If you wanted to buy Dolby-grade voice cleaning for a live conferencing product, that was the door.

In 2024 that door began to close. In August 2024 Dolby acquired THEO Technologies, a maker of low-latency streaming players, and through 2025–2026 folded its developer platform into a new brand, Dolby OptiView. OptiView's products today are a video player, live and real-time streaming, and advertising tools — there is no self-service real-time conferencing-and-voice product among them. The old Communications documentation now redirects into the OptiView streaming docs. Dolby still makes excellent audio technology, and a separate file-based media-enhancement API for processing recorded audio exists, but the live, self-service "buy Dolby Voice for your call" path that competing articles still describe has effectively been wound down.

That is the lesson, and it is worth more than any feature comparison: a "buy" option is a bet on a vendor's roadmap, and roadmaps move. A team that built its conferencing product on Dolby's Communications APIs in 2023 spent 2024–2026 migrating off them. The risk is not that Dolby did anything wrong — companies refocus — but that outsourcing a core capability means inheriting someone else's strategy. When you buy, you must read not just the demo but the direction.

Figure 2. Dolby's developer platform moved from self-service real-time Communications (with Dolby Voice) to streaming-focused Dolby OptiView after the 2024 THEO acquisition — the clearest reminder that a bought capability rides on a vendor's roadmap.

A Common Mistake: Buying The Demo Instead Of The Fit

The most expensive error in this whole topic is choosing a vendor from a sound demo. Every engine on this page sounds impressive in a controlled clip; the demo is the part that is easy. The questions that actually decide success are never in the demo, and skipping them is how teams end up rebuilding a year later.

Ask these before any contract. Does the engine run where your audio physically is? Krisp runs on the user's device, so it is useless for cleaning audio that arrives from a phone line you do not control; Maxine runs on your GPUs, so it is useless if you have no GPUs and do not want a fleet. Does the vendor's roadmap point the same way as your product, or are you about to become the next Dolby migration story? Does the price scale with the thing your product scales with — concurrent streams, devices, or contract seats? And can you verify the quality with a real, repeatable test rather than the vendor's clip — for noise suppressors that means an ITU-T P.835 style listening evaluation, where humans score the speech, the background, and the overall result separately, so you measure improvement honestly instead of trusting a marketing waveform. A vendor that passes the demo but fails any of these four questions is the wrong buy.

The Vendors Side By Side

With each option placed on the spectrum and the demo trap named, the comparison table becomes readable as operations rather than a leaderboard.

Read the table as a set of trade-offs, not winners. Build is the cheapest licence and the most expensive team. Krisp is the lowest operational burden — nothing runs on your servers — at the price of an enterprise contract. Maxine is the highest ceiling and the widest effect set, paid for in GPUs and operations. Dolby is the reminder to check the roadmap before you commit. There is also a fast-growing fourth shape worth knowing: newer managed vendors such as ai-coustics offer an on-device, CPU-only SDK — its real-time model runs in under thirty milliseconds with a small footprint and no GPU — sold with self-service, pay-as-you-grow pricing rather than an upfront enterprise contract, which narrows Krisp's old advantage of being the only turnkey option.

From Spectrum To Decision

Turn all of this into a path you can actually walk. The decision tree below settles, in order, the questions that matter: whether the free baseline is enough, whether your audio comes from devices you control, and whether you already operate GPUs — because each answer eliminates options and points at one home.

Figure 3. A decision path: try the free baseline first, then choose by where your audio comes from and whether you already operate GPUs — not by which demo sounds best.

The tree encodes a simple priority. Start free: if the browser's built-in suppressor already silences the complaints, ship it and spend your budget elsewhere. If not, the deciding question is where your audio comes from. If some of it arrives from phones, AI agents, or devices you do not control, you need a server-side engine, and then the question is whether you already run NVIDIA GPUs — if you do, Maxine's effect suite is a natural fit; if you do not, a managed on-device SDK that you also run server-side, or a CPU-only managed engine, avoids standing up a GPU fleet. If every participant is in a browser or app you control, you choose between buying a turnkey on-device SDK (Krisp or ai-coustics) for zero server operations, or building with RNNoise or DeepFilterNet for a free licence and full control. One pitfall to avoid throughout: never run two suppressors at once — the built-in one plus a bought one double-processes the audio and makes it worse, a trap covered in detail in the WebRTC integration article.

Why The Comparison Keeps Shifting

One reason this topic is hard is that the ground keeps moving, and the three names moved in three different directions in just two years. Krisp expanded from noise cancellation into a broader voice-AI toolkit for agents — turn-taking, voice isolation, accent conversion — chasing the boom in AI voice agents. NVIDIA rebranded Maxine to AI for Media and pushed it deeper into broadcast and GPU-served pipelines. Dolby exited self-service real-time voice entirely and refocused on streaming. A comparison written in 2023 would get all three wrong today. The practical defence is to treat your vendor choice as a decision with a review date, not a permanent marriage — re-confirm, at least yearly, that the engine you bought still runs where you need it, still costs what you expected, and still sits on a vendor roadmap pointed your way. Echo cancellation, a neighbouring problem these engines also touch, has its own moving landscape covered in our echo cancellation in WebRTC article.

Where Fora Soft Fits In

We build the live-video products that depend on clean audio — video conferencing platforms, telemedicine consults where a doctor must catch every word, e-learning classrooms, and AI meeting tools — and in that work the build-versus-buy call comes up on almost every project. The pattern we apply is the spectrum above, settled in order: start with the free built-in baseline, reach for an open model like RNNoise when a browser-only product needs more and the team wants control, and move to a managed engine when a product spans many platforms, takes phone-in callers, or runs AI voice agents on devices we do not control. For privacy-sensitive verticals like telemedicine, an on-device engine that uploads no audio is often the deciding factor by itself. The discipline that saves the most rework is the one this article argues for: choose by where the audio lives and how the cost scales, verify with a real listening test rather than a vendor clip, and put a yearly review on the vendor's roadmap so a shift like Dolby's never strands a shipped product.

What To Read Next

• Real-time noise suppression in WebRTC — RNNoise, DeepFilterNet, and Krisp integration
• Real-time noise suppression models — RNNoise, DeepFilterNet, and Krisp compared
• The real cost of AI in video products — Gemini, OpenAI, and GPU pricing

Talk To Us / See Our Work / Download

• Talk to a video engineer about choosing and integrating a voice-enhancement engine for your product → /services/webrtc-development
• See our case studies in video conferencing, telemedicine, and e-learning → /cases
• Download the Voice-Enhancement Build-vs-Buy Decision Sheet (one page, printable) → Download the decision sheet

References

1. Krisp. Real-Time AI Voice SDK and Developer Hub (product and developer documentation), accessed 2026-06-02. https://krisp.ai/developers/. First-party source for Krisp's on-device processing (audio never uploaded), platform coverage (Windows, macOS, Linux, Android, iOS, browser via WebAssembly), the SDK capability set (Noise Cancellation inbound/outbound, Background Voice Cancellation, Voice Isolation for voice agents, Accent Conversion, per-frame statistics), the ~25 ms (≈15 ms lite) added latency, the 200M+ device / 75B+ minutes-per-month deployment scale, and the Discord / RingCentral / Zoho / Twilio / Daily deployment base.
2. NVIDIA. NVIDIA AI for Media (formerly NVIDIA Maxine) product page, accessed 2026-06-02. https://developer.nvidia.com/maxine. First-party source for the 2026 rename from Maxine to AI for Media, the Audio Effects component (noise removal, room-echo removal, audio super-resolution, acoustic echo cancellation), the Studio Voice model, NIM-microservice and SDK delivery, and the free API-catalogue prototyping plus 90-day evaluation-licence path.
3. NVIDIA. Audio Effects (AFX) SDK — User Guide and Programming Guide, accessed 2026-06-02. https://docs.nvidia.com/maxine/afx/latest/index.html. First-party source for the GPU requirement (NVIDIA GPU with Tensor Cores, minimum 10 GB RAM), the Tensor-Core acceleration of the audio algorithms, and the combined noise-plus-room-echo removal behaviour.
4. NVIDIA. NVIDIA AI Enterprise (product and licensing), accessed 2026-06-02. https://www.nvidia.com/en-us/data-center/products/ai-enterprise/. Source for the production licensing of Maxine / AI for Media NIM microservices through NVIDIA AI Enterprise — the ~$4,500 per GPU per year subscription (≈$1 per GPU-hour in the cloud) used in the cost example, on top of GPU hardware or rental.
5. Dolby. Dolby OptiView Documentation, accessed 2026-06-02. https://optiview.dolby.com/docs/. First-party source establishing that Dolby's developer platform is now Dolby OptiView — products are the OptiView Player (formerly THEOplayer), Live and Real-time streaming (THEOlive, Millicast), Ads, and Ad Engine — with no self-service real-time Communications / voice product; the former Communications documentation URL redirects here. Basis for the "the real-time voice buy path has closed" claim.
6. Dolby. THEO Technologies is joining Dolby (newsroom announcement), 1 August 2024, accessed 2026-06-02. https://news.dolby.com/en-WW/240056-theo-technologies-is-joining-dolby/. First-party source for the August 2024 acquisition of THEO Technologies (low-latency streaming players) and the subsequent refocus of the developer platform toward streaming under Dolby OptiView.
7. ai-coustics. Comparing Krisp and ai-coustics real-time audio enhancement (vendor blog), 10 November 2025, accessed 2026-06-02. https://ai-coustics.com/blog/comparing-krisp-and-ai-coustics-real-time-audio-enhancement-which-is-best-for-you. Competitor source (treated as a competitor's framing, not a neutral fact) for the characterisation of Krisp's SDK pricing as an upfront annual enterprise licence, and the first-party claims about ai-coustics' Quail real-time model: <30 ms latency, <10 MB runtime, CPU-only (no GPU), on-device, self-service licensing.
8. LiveKit. Noise & echo cancellation (documentation), accessed 2026-06-02. https://docs.livekit.io/transport/media/noise-cancellation/. Vendor source for how managed engines (Krisp and ai-coustics) attach to a real-time platform, the noise-cancellation-versus-background-voice-cancellation distinction, the do-not-double-process rule, and the published word-error-rate comparison (raw 117.6%, Krisp BVC 23.5%, ai-coustics Voice Focus 7.1%) used to anchor the "verify with a real test" point.
9. ITU-T. Recommendation P.835 — Subjective test methodology for evaluating speech communication systems that include noise suppression algorithm, approved 13 November 2003, accessed 2026-06-02. https://www.itu.int/rec/T-REC-P.835. Primary standards source for the three-axis (SIG / BAK / OVRL) human evaluation of noise suppressors — the honest way to compare engines beyond a vendor demo.
10. ITU-T. Recommendation G.114 — One-way transmission time, accessed 2026-06-02. https://www.itu.int/rec/T-REC-G.114. Primary standards source for the ≤150 ms one-way mouth-to-ear delay budget that any bought engine's added latency (e.g. Krisp's ~25 ms) must fit inside.
11. IETF. RFC 7874 — WebRTC Audio Codec and Processing Requirements, May 2016, accessed 2026-06-02. https://www.rfc-editor.org/rfc/rfc7874. Primary standards source for the mandatory WebRTC audio codecs (Opus, G.711) and processing context — the basis for the rule that voice enhancement must run on raw audio before the encoder, which constrains where each vendor's engine can attach.
12. W3C. Media Capture and Streams (Recommendation), accessed 2026-06-02. https://www.w3.org/TR/mediacapture-streams/. Primary standards source for getUserMedia and the noiseSuppression constrainable property — the free built-in baseline the decision tree tries first, and the switch you must disable to avoid double-processing with a bought engine.
13. W3C. Web Audio API (Recommendation, 17 June 2021), accessed 2026-06-02. https://www.w3.org/TR/webaudio/. Primary standards source for AudioWorklet, the off-main-thread mechanism that hosts a built or bought on-device suppressor on raw audio before the encoder.
14. getLatka. Krisp revenue and company profile, 2025/2026, accessed 2026-06-02. https://getlatka.com/companies/krisp.ai. Secondary source for Krisp company context only (founded 2017, Berkeley headquarters, ~$37.7M reported 2025 revenue, ~343-person team) — not used for any technical claim.