Hush Review (2026): Pricing, Features & Honest Verdict

TLDR: Hush by weya AI is an open-source speech enhancement model that isolates the main speaker and strips out background noise in real time, all from an 8 MB file that runs on a CPU. We think it’s a genuinely impressive gift to voice-AI builders, with the catch that it’s a model you wire in yourself, not a finished app. Best for: developers building voice agents, call-center AI, or transcription who want clean input audio on their own infrastructure. Price: Free, open source (free plan: yes). Rating: 8/10.

What Is Hush?

Hush is an open-source speech enhancement model built by weya AI, an Indian company working on an on-premises voice AI stack for the banking and financial sector. We want to flag the name early, because there are unrelated apps called Hush floating around. This one is a developer tool, not a consumer app. Its job is narrow and useful: take a live audio stream, lock onto the primary speaker, and suppress everything else. That means background noise, secondary speakers talking over each other, plus the annoying stuff like hum, hiss, and whistles. The output is cleaner audio for whatever comes next, usually a voice agent or a transcription model. We tested the premise against the kind of audio that wrecks most pipelines, the cafe background hum and the second person talking over the first, and that’s exactly the mess Hush is tuned to handle.

The reason that matters is that voice AI tends to fall apart in the real world. A model that transcribes beautifully in a quiet room starts hallucinating words the moment someone’s in a call center, a car, or a busy office. Hush sits at the front of the pipeline as a cleanup layer. weya AI released it in June 2026 and put the full weights and source on Hugging Face and GitHub under Apache 2.0. At launch it ranked #5 on Hugging Face’s Audio-to-Audio leaderboard, which makes it one of the stronger open-source options in its category. And the company isn’t a weekend project. weya AI counts Tier-1 institutions like Kotak Bank among its clients, which is more credibility than most open models arrive with.

What Are Hush’s Key Features?

Real-time speaker isolation and noise suppression

The core feature is exactly what you’d want from a front-end audio model. Hush isolates the primary speaker from a live stream and suppresses competing voices, background noise, secondary speakers, and tonal junk like hum, hiss, and whistles. It does this in real time, not as a batch post-process. That distinction is the whole point. For a voice agent to feel responsive, the cleanup has to happen as the audio arrives, and Hush is designed for that streaming case rather than cleaning up a recording after the fact.

Tiny footprint and CPU-only

This is the part that made us look twice. Hush is 8 MB. It has 1.8 million parameters, runs with no GPU, and processes audio in under 1 ms per 10 ms frame. In plain terms, it keeps up with live audio comfortably and leaves you headroom to spare. You can run it at the edge or on-device without a graphics card, which is a big deal for anyone who doesn’t want to pay for GPU instances just to clean up microphone input. Most of the polished alternatives assume a hosted service or a beefier setup. Hush fits on a Raspberry Pi. And because the file is so small, shipping it inside an existing app or container adds almost nothing to your build size, which is the kind of practical detail that makes engineers actually adopt a model rather than just star it on GitHub.

Language-agnostic by design

Hush was trained on over 10,000 hours of mixed data and is fully language-agnostic. It works on the acoustic properties of speech rather than the words being spoken, so it doesn’t care whether your users are speaking English, Hindi, Spanish, or anything else. For a company like weya AI serving Indian banking customers, that’s not a marketing line, it’s a requirement. And it’s useful for anyone shipping a product across multiple regions who doesn’t want to maintain a separate cleanup model per language.

Fully open source, no lock-in

The model weights and complete source are on Hugging Face and GitHub under the Apache 2.0 license. That’s the permissive kind. You can self-host it, modify it, and ship it in a commercial product without asking permission or paying a per-minute fee. There’s no API call leaving your network, which matters a lot if you’re handling sensitive audio like bank calls. You own the whole thing once you’ve integrated it.

How Much Does Hush Cost?

Hush is free. The model is released under Apache 2.0, so the weights and source cost you nothing and there are no usage fees, no per-minute billing, and no seat licenses. You download it and run it on your own hardware.

The honest catch is that “free” here means free as in code, not free as in effort. The real cost of Hush is engineering. Someone on your team has to integrate it into your audio pipeline, test it against your actual conditions, and maintain it. There’s no hosted endpoint you can point at and no support contract to buy for the model itself. weya AI makes its money elsewhere: it sells a broader on-premises voice AI stack and enterprise deployment work to institutions like banks. Hush is the open, free piece that demonstrates the company’s chops and feeds the top of that funnel.

Compared with the alternatives, that’s a meaningful price gap. Krisp is polished and hosted but it’s a paid product across its consumer and enterprise tiers. Picovoice Koala is an on-device SDK with commercial licensing, so you’re paying to use it in production. NVIDIA Maxine needs GPUs. Hush asks for none of that money, only your developers’ time. For a team already comfortable running models in their own infrastructure, that trade usually favors Hush.

Who Is Hush Best For?

Use Hush if you’re a developer or team building voice agents, call-center AI, transcription, or any real-time voice product, and you want a clean audio front-end you run yourself. It’s a strong fit if you care about keeping audio on your own infrastructure for privacy or compliance reasons, if you want to avoid per-minute API costs at scale, and if you have the engineering capacity to integrate and maintain a model. The CPU-only footprint makes it especially attractive for edge and on-device deployments where a GPU isn’t an option.

Skip Hush if you’re non-technical and want a plug-and-play app you can install and forget. This is a model, not a product, and there’s no friendly desktop client. Skip it too if you need a managed, hosted API with a support SLA out of the box, because Hush gives you neither for the model itself. And skip it if your problem is anything other than speech. Hush is built for isolating a human voice. It’s the wrong tool for cleaning up music or general audio.

Best Hush Alternatives

Krisp. The polished, hosted option. Krisp does noise and voice cancellation across consumer and enterprise products and it just works out of the box, with apps and SDKs and support. The trade-off is that it’s paid, and you’re routing through their stack rather than owning the model. If you want zero engineering effort and don’t mind the bill, Krisp is the easy pick. If you want to self-host and keep audio in-house, Hush wins.

Picovoice Koala. An on-device noise suppression SDK with commercial licensing. Koala is closer to Hush in spirit, since it runs locally and targets edge use cases. The difference is the license: Koala is a commercial product you pay to ship, while Hush is Apache 2.0 and free. If you want vendor support and a packaged SDK, Koala has the edge. If cost and openness matter more, Hush does.

RNNoise. The classic open-source noise suppression library. It’s free like Hush and has been around for years, which makes it battle-tested and widely embedded. But it’s older and less capable, and it’s about noise reduction rather than the speaker isolation Hush does. If you only need basic noise cleanup and want something proven, RNNoise is fine. For competing-voice suppression and leaderboard-level quality, Hush is a clear step up.

Final Verdict: Is Hush Worth It?

For what it is, Hush is excellent. An 8 MB model that runs on a CPU, processes live audio with sub-millisecond latency per frame, works across languages, and ranks near the top of an open leaderboard is a real gift to anyone building voice products. The Apache 2.0 license seals it. You can ship this in production without paying weya AI a cent, and the fact that weya AI is a real company with banking clients tells you the model came out of solving actual problems, not a demo.

The honest reservations are about fit, not quality. This is a model you integrate, so budget for engineering time and accept that there’s no hosted API or official support tier to lean on. It’s also speech-only. If those things are dealbreakers, look at Krisp instead. But if you have developers, want clean audio without per-minute fees, and don’t want to send sensitive voice data to a third party, Hush deserves a serious look. We’d recommend it to almost any team building real-time voice AI on their own infrastructure. It earns an 8/10.