{"slug": "introducing-voice-isolator", "title": "Introducing Voice Isolator", "summary": "Inworld AI released Voice Isolator, its first open-source model, under the Apache 2.0 license with code and weights on GitHub and Hugging Face. Voice Isolator runs at 48 kHz and ships as two checkpoints: an offline model for recorded audio and a realtime streaming model with 20 ms of look-ahead, both trained to preserve speech, laughter and breaths while removing background chatter, music and transients. Inworld AI said its models scored best among the open-source denoisers it tested on VIBE (Voice Isolation Benchmark and Evals), a private multilingual benchmark the company built and has not yet published.", "body_md": "# Introducing *Voice Isolator*\n\nRemove background noise while keeping speech, laughter and breaths. Two 48 kHz models, for offline processing and for realtime streaming, with code and weights under Apache 2.0.[View on GitHub ↗](https://github.com/inworld-ai/voice-isolator)[Hugging Face ↗](https://huggingface.co/inworld/voice-isolator)\n\nToday we're releasing our first open-source model: **Voice Isolator**. It keeps the voice as it was recorded, with its timbre, breaths and level, and removes background chatter, music and transients. It works at 48 kHz and ships as two checkpoints: an offline model for the best quality on existing recordings, and a realtime model that streams with 20 ms of look-ahead for live applications.\n\n## 01Purpose\n\nRemoving traffic, music or someone talking in the background is half the job. The other half is keeping the person you want to hear. Most denoisers treat a laugh, a breath or a sigh as noise and take it out. Voice Isolator is trained to keep them.\n\nListen to it on four noisy clips. Switch isolation off to hear what the microphone captured.\n\nIn a live conversation, those sounds tell a voice agent a lot: a breath before someone speaks, a laugh in response, or an \"hmm\" while they think. Keeping them helps the agent decide [when to respond](https://inworld.ai/blog/introducing-sava). In a recorded interview or a voice sample, they keep the character of the original voice.\n\n## 02Evals\n\nWe built VIBE (Voice Isolation Benchmark and Evals) to measure how well a model preserves the original voice while removing everything else. It is a multilingual, multi-condition set of real utterances, with and without added noise, and we use it to score open-source models against each other.\n\nOur models performed best among the open-source denoisers we were able to test.\n\nVIBE is a private benchmark for now. If you'd like to submit a model, or help improve Voice Isolator or the benchmark itself, [reach out](https://github.com/inworld-ai/voice-isolator/issues) and we'll be happy to collaborate.\n\n## 03How to\n\nStart with a recording you know well. Install the package, then run it on the file using the commands below. The weights download from [Hugging Face](https://huggingface.co/inworld/voice-isolator) on first use. The model runs on CUDA, Apple silicon and CPUs.\n\n```\npip install inworld-voice-isolation\n\n# the voice goes to clean.wav, what was removed to clean_background.wav\nvoice-isolate noisy.wav clean.wav\n\n# the realtime model, as a stream\nvoice-isolate noisy.wav clean.wav --mode streaming --lookahead 2\npython\nfrom inworld import voice_isolation\n\nlive = voice_isolation.VoiceIsolator(\"inworld/voice-isolator\",\n                                     mode=\"streaming\", lookahead_frames=2)\nsession = live.stream()  # 48 kHz mono, in pieces of any size\nfor piece in microphone:\n    play(session.push(piece))  # the samples that are final\nplay(session.flush())\n```\n\nTo hear it for yourself, run the repository's [interactive demo](https://github.com/inworld-ai/voice-isolator/blob/main/demo/README.md) on your laptop. It has a tab for recorded files and another that sends microphone audio through the realtime model.\n\n## 04Next\n\nVoice Isolator currently only does denoising. We trained it this way on purpose, so that it introduces no bias into downstream training tasks or into realtime speech-to-text models, which can be sensitive to changes in the waveform.\n\nBut denoising alone won't fix a low-quality microphone or missing band detail. That is restoration. We're looking at it, and we may have something to share.\n\nWe'd like to hear how Voice Isolator works in your application. If you have an idea for making it better, or you found something that doesn't work well, create an [issue](https://github.com/inworld-ai/voice-isolator/issues) or ask your agent to open a pull request, and we'll respond. Let's make it better together.", "url": "https://wpnews.pro/news/introducing-voice-isolator", "canonical_source": "https://inworld.ai/blog/introducing-voice-isolator", "published_at": "2026-11-18 00:00:00+00:00", "updated_at": "2026-11-19 09:43:59.750549+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-tools", "developer-tools", "ai-agents"], "entities": ["Inworld AI", "Voice Isolator", "VIBE", "Hugging Face", "GitHub"], "also_reported_by": [], "takeaway": "Voice Isolator ships two Apache 2.0 checkpoints at 48 kHz — an offline model and a streaming model with 20 ms of look-ahead — installable via `pip install inworld-voice-isolation` and runnable on CUDA, Apple silicon and CPUs.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/introducing-voice-isolator", "markdown": "https://wpnews.pro/news/introducing-voice-isolator.md", "text": "https://wpnews.pro/news/introducing-voice-isolator.txt", "jsonld": "https://wpnews.pro/news/introducing-voice-isolator.jsonld"}}