Artificial intelligence news.
Straight from the original announcements.

Inworld AI Releases Open-Source Voice Isolator Model

November 18, 2026By the AI News Bites Editorial TeamSource: Inworld AI
Inworld AI Releases Open-Source Voice Isolator Model
Image: Inworld AI

Inworld AI has released Voice Isolator, its first open-source model, which removes background chatter, music and other transients while preserving the recorded voice. The model runs at 48 kHz and includes separate offline and realtime checkpoints.

Key points
  • Voice Isolator is Inworld AI’s first open-source model.
  • It preserves timbre, breaths and other voice details while removing background audio.
  • The release includes offline and realtime 48 kHz checkpoints.
  • The model runs on CUDA, Apple silicon and CPUs.
  • Inworld AI reports VIBE scores of 0.768 offline and 0.718 for streaming.

Designed to preserve the voice, not just remove noise

Voice Isolator is designed to retain the timbre, breaths and level of a voice while reducing unwanted sounds such as traffic, music, other speakers and household noise. Inworld AI says the model was trained to keep details that conventional denoisers can remove, including laughs, breaths and sighs.

That distinction matters for both voice agents and recorded audio. In a live conversation, a breath before speech, a laugh or an "hmm" can help an agent decide when to respond. For interviews and voice samples, those sounds help preserve the character of the original recording.

The release includes two checkpoints. The offline model is intended to provide the best quality on existing recordings, while the realtime model streams audio with 20 ms of look-ahead for live applications.

VIBE benchmark compares open-source denoisers

Inworld AI also introduced VIBE, which stands for Voice Isolation Benchmark and Evals. The company describes it as a multilingual, multi-condition collection of real utterances, with and without added noise, used to measure how well a model preserves the original voice while removing other audio.

On the company’s VIBE results, Voice Isolator scored 0.768, while its realtime streaming version scored 0.718. The listed scores for other open-source denoisers were 0.695 for ClearerVoice MossFormer2_SE_48K, 0.646 for resemble-enhance (denoise), 0.589 for FastEnhancer-L 48k Streaming and 0.510 for DeepFilterNet3 Streaming.

VIBE scores range from 0, representing the unprocessed input, to 1, representing the clean reference. According to Inworld AI, its models performed best among the open-source denoisers it tested. The company says the benchmark is private for now, although it is inviting model submissions and contributions.

Open-source package supports offline and streaming use

Developers can install the package with pip install inworld-voice-isolation and run Voice Isolator on an audio file. The tool produces a cleaned voice file and a separate file containing the removed background audio. The weights download from Hugging Face on first use.

The model runs on CUDA, Apple silicon and CPUs. The streaming interface accepts 48 kHz mono audio in pieces of any size, and the repository includes an interactive laptop demo with separate tabs for recorded files and microphone input.

Inworld AI says Voice Isolator currently performs denoising only. The company deliberately trained it this way to avoid introducing bias into downstream training tasks and realtime speech-to-text models, which can be sensitive to waveform changes. It is considering restoration work for low-quality microphones and missing band detail, but says denoising alone does not address those problems.

Source
Inworld AI: Introducing Voice Isolator

Related news

Hugging Face highlights Rumble-2-Flash, an open 0.5B speech model
ModelsNovember 18, 2026

Hugging Face highlights Rumble-2-Flash, an open 0.5B speech model

Hugging Face is highlighting Rumble-2-Flash, an open automatic speech recognition model with 0.5B parameters and support for more than 40 languages. The model is…

Inworld releases SAVA model for faster voice-agent turn detection
ModelsNovember 18, 2026

Inworld releases SAVA model for faster voice-agent turn detection

Inworld AI has released SAVA, a semantic acoustic model that combines speech activity detection with turn detection for voice applications. The company says the model…

Egyptian and Nigerian English Added to Hugging Face ASR Leaderboard
ResearchNovember 6, 2026

Egyptian and Nigerian English Added to Hugging Face ASR Leaderboard

Egyptian and Nigerian English have joined Hugging Face’s Open ASR Leaderboard, expanding the English varieties represented on the speech recognition evaluation platform.…